Julio Marchi © Speaks Out

New Post has been published on https://thedigitalinsider.com/google-reveals-its-own-version-of-apples-ai-cloud/

Google reveals its own version of Apple’s AI cloud

Google has rolled out Private AI Compute, a new cloud-based processing system designed to bring the privacy of on-device AI to the cloud. The platform aims to give users faster, more capable AI experiences without compromising data security. It combines Google’s most advanced Gemini models with strict privacy safeguards, reflecting the company’s ongoing effort to make AI both powerful and responsible.

The feature closely resembles Apple’s Private Cloud Compute, signalling how major tech firms are rethinking privacy in the age of large-scale AI. Both companies are trying to balance two competing needs — the huge computing power required to run advanced AI models and users’ expectations for data privacy.

Why Google built Private AI Compute

As AI systems get smarter, they’re also becoming more personal. What started as tools that completed simple tasks or answered direct questions are now systems that can anticipate user needs, suggest actions, and handle complex processes in real time. That kind of intelligence demands a level of reasoning and computation that often exceeds what’s possible on a single device.

Private AI Compute bridges that gap. It lets Gemini models in the cloud process data faster and more efficiently while ensuring that sensitive information remains private and inaccessible to anyone else — not even Google engineers. Google describes it as combining the power of cloud AI with the security users expect from local processing.

In practical terms, this means you could get quicker responses, smarter suggestions, and more personalised results without your personal data ever leaving your control.

How Private AI Compute keeps data secure

Google claims the new platform is based on the same principles that underpin its broader AI and privacy strategy: giving users control, maintaining security, and earning trust. The system acts as a protected computing environment, isolating data so it can be processed safely and privately.

It uses a multi-layered design centred on three key components:

Unified Google tech stack: Private AI Compute runs entirely on Google’s own infrastructure, powered by custom Tensor Processing Units (TPUs). It’s secured through Titanium Intelligence Enclaves (TIE), which create an additional layer of protection for data processed in the cloud.

Encrypted connections: Before data is sent for processing, remote attestation and encryption verify that it’s connecting to a trusted, hardware-secured environment. Once inside this sealed cloud space, information stays private to the user.

Zero access assurance: Google says the system is designed so that no one — not even the company itself — can access the data processed within Private AI Compute.

This design builds on Google’s Secure AI Framework (SAIF), AI Principles, and Privacy Principles, which outline how the company develops and deploys AI responsibly.

What users can expect

Private AI Compute also improves the performance of AI features that are already running on devices. Magic Cue on the Pixel 10 can now offer more relevant and timely suggestions by leveraging cloud-level processing power. Similarly, the Recorder app can use the system to summarise transcriptions across a wider range of languages — something that would be difficult to do entirely on-device.

These examples hint at what’s ahead. With Private AI Compute, Google can deliver AI experiences that combine the privacy of local models with the intelligence of cloud-based ones. It’s an approach that could eventually apply to everything from personal assistants and photo organisation to productivity and accessibility tools.

Google calls this launch “just the beginning.” The company says Private AI Compute opens the door to a new generation of AI tools that are both more capable and more private. As AI becomes increasingly woven into everyday tasks, users are demanding greater transparency and control over how their data is used — and Google appears to be positioning this technology as part of that answer.

For those interested in the technical details, Google has published a technical brief explaining how Private AI Compute works and how it fits into the company’s larger vision for responsible AI development.

(Photo by Solen Feyissa)

See also: Apple plans big Siri update with help from Google AI

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

New Post has been published on https://thedigitalinsider.com/wiz-security-lapses-emerge-amid-the-global-ai-race/

Wiz: Security lapses emerge amid the global AI race

According to Wiz, the race among AI companies is causing many to overlook basic security hygiene practices.

65 percent of the 50 leading AI firms the cybersecurity firm analysed had leaked verified secrets on GitHub. The exposures include API keys, tokens, and sensitive credentials, often buried in code repositories that standard security tools do not check.

Glyn Morgan, Country Manager for UK&I at Salt Security, described this trend as a preventable and basic error. “When AI firms accidentally expose their API keys they lay bare a glaring avoidable security failure,” he said.

“It’s the textbook example of governance paired with a security configuration, two of the risk categories that OWASP flags. By pushing credentials into code repositories they hand attackers a golden ticket to systems, data, and models, effectively sidestepping the usual defensive layers.”

Wiz’s report highlights the increasingly complex supply chain security risk. The problem extends beyond internal development teams; as enterprises increasingly partner with AI startups, they may inherit their security posture. The researchers warn that some of the leaks they found “could have exposed organisational structures, training data, or even private models.”

The financial stakes are considerable. The companies analysed with verified leaks have a combined valuation of over $400 billion.

The report, which focused on companies listed in the Forbes AI 50, provides examples of the risks:

LangChain was found to have exposed multiple Langsmith API keys, some with permissions to manage the organisation and list its members. This type of information is highly valued by attackers for reconnaissance.

An enterprise-tier API key for ElevenLabs was discovered sitting in a plaintext file.

An unnamed AI 50 company had a HuggingFace token exposed in a deleted code fork. This single token “allow[ed] access to about 1K private models”. The same company also leaked WeightsAndBiases keys, exposing the “training data for many private models.”

The Wiz report suggests this problem is so prevalent because traditional security scanning methods are no longer sufficient. Relying on basic scans of a company’s main GitHub repositories is a “commoditised approach” that misses the most severe risks .

The researchers describe the situation as an “iceberg” (i.e. the most obvious risks are visible, but the greater danger lies “below the surface”.) To find these hidden risks, the researchers adopted a three-dimensional scanning methodology they call “Depth, Perimeter, and Coverage”:

Depth: Their deep scan analysed the “full commit history, commit history on forks, deleted forks, workflow logs and gists”—areas most scanners “never touch”.

Perimeter: The scan was expanded beyond the core company organisation to include organisation members and contributors. These individuals might “inadvertently check company-related secrets into their own public repositories”. The team identified these adjacent accounts by tracking code contributors, organisation followers, and even “correlations in related networks like HuggingFace and npm.”

Coverage: The researchers specifically looked for new AI-related secret types that traditional scanners often miss, such as keys for platforms like WeightsAndBiases, Groq, and Perplexity.

This expanded attack surface is particularly worrying given the apparent lack of security maturity at many fast-moving companies. The report notes that when researchers tried to disclose the leaks, almost half of disclosures either failed to reach the target or received no response. Many firms lacked an official disclosure channel or simply failed to resolve the issue when notified.

Wiz’s findings serve as a warning for enterprise technology executives, highlighting three immediate action items for managing both internal and third-party security risk.

Security leaders must treat their employees as part of their company’s attack surface. The report recommends creating a Version Control System (VCS) member policy to be applied during employee onboarding. This policy should mandate practices such as using multi-factor authentication for personal accounts and maintaining a strict separation between personal and professional activity on platforms like GitHub.

Internal secret scanning must evolve beyond basic repository checks. The report urges companies to mandate public VCS secret scanning as a “non-negotiable defense”. This scanning must adopt the aforementioned “Depth, Perimeter, and Coverage” mindset to find threats lurking below the surface.

This level of scrutiny must be extended to the entire AI supply chain. When evaluating or integrating tools from AI vendors, CISOs should probe their secrets management and vulnerability disclosure practices. The report notes that many AI service providers are leaking their own API keys and should “prioritise detection for their own secret types.”

The central message for enterprises is that the tools and platforms defining the next generation of technology are being built at a pace that often outstrips security governance. As Wiz concludes, “For AI innovators, the message is clear: speed cannot compromise security”. For the enterprises that depend on that innovation, the same warning applies.

See also: Exclusive: Dubai’s Digital Government chief says speed trumps spending in AI efficiency race

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security Expo, click here for more information.

New Post has been published on https://thedigitalinsider.com/black-friday-deal-save-big-on-avid-media-composer/

Black Friday Deal: Save Big on Avid Media Composer!

Power and Precision for Every Editor Avid Media Composer | Ultimate gives professionals the speed, precision, and creative tools to tell their best stories. Trusted by top film and TV editors, it combines ACE-certified tools with reliable media management to streamline every stage of production. Work faster and with complete confidence in every cut.

Seamless Collaboration Anywhere Media Composer | Ultimate makes teamwork easy. Share bins, projects, and media with editors and producers locally or remotely through Avid NEXIS and MediaCentral. Collaborate across platforms—including Premiere Pro and Final Cut Pro—and keep everything in sync with Cloud Remote support

New Post has been published on https://thedigitalinsider.com/chinese-ai-startup-moonshot-outperforms-gpt-5-and-claude-sonnet-4-5-what-you-need-to-know/

Chinese AI startup Moonshot outperforms GPT-5 and Claude Sonnet 4.5: What you need to know

A Chinese AI startup, Moonshot, has disrupted expectations in artificial intelligence development after its Kimi K2 Thinking model surpassed OpenAI’s GPT-5 and Anthropic’s Claude Sonnet 4.5 across multiple performance benchmarks, sparking renewed debate about whether America’s AI dominance is being challenged by cost-efficient Chinese innovation.

Beijing-based Moonshot AI, valued at US$3.3 billion and backed by tech giants Alibaba Group Holding and Tencent Holdings, released the open-source Kimi K2 Thinking model on November 6, achieving what industry observers are calling another “DeepSeek moment” – a reference to the Hangzhou-based startup’s earlier disruption of AI cost assumptions.

🚀 Hello, Kimi K2 Thinking! The Open-Source Thinking Agent Model is here.

🔹 SOTA on HLE (44.9%) and BrowseComp (60.2%) 🔹 Executes up to 200 – 300 sequential tool calls without human interference 🔹 Excels in reasoning, agentic search, and coding 🔹 256K context window

Built… pic.twitter.com/lZCNBIgbV2

— Kimi.ai (@Kimi_Moonshot) November 6, 2025

Performance metrics challenge US models

According to the company’s GitHub blog post, Kimi K2 Thinking scored 44.9% on Humanity’s Last Exam, a large language model benchmark consisting of 2,500 questions across a broad range of subjects, exceeding GPT-5’s 41.7%.

The model also achieved 60.2% on the BrowseComp benchmark, which evaluates web browsing proficiency and information-seeking persistence of large language model agents, and scored 56.3% to lead in the Seal-0 benchmark designed to challenge search-augmented models on real-world research queries.

VentureBeat reported that the fully open-weight release meeting or exceeding GPT-5’s scores marks a turning point where the gap between closed frontier systems and publicly available models has effectively collapsed for high-end reasoning and coding.

Kimi K2 Thinking is the new leading open weights model: it demonstrates particular strength in agentic contexts but is very verbose, generating the most tokens of any model in completing our Intelligence Index evals@Kimi_Moonshot‘s Kimi K2 Thinking achieves a 67 in the… pic.twitter.com/m6SvpW7iif

— Artificial Analysis (@ArtificialAnlys) November 7, 2025

Cost efficiency raises questions

The popularity of the model grew after CNBC reported its training cost was merely US$4.6 million, though Moonshot AI did not comment on the cost. According to calculations by the South China Morning Post, the cost of Kimi K2 Thinking’s application programming interface was six to 10 times cheaper than that of OpenAI and Anthropic’s models.

The model uses a Mixture-of-Experts architecture with one trillion total parameters, of which 32 billion are activated per inference, and was trained using INT4 quantisation to achieve roughly two times generation speed improvement while maintaining state-of-the-art performance.

Thomas Wolf, co-founder of Hugging Face, commented on X that Kimi K2 Thinking was another case of an open-source model passing a closed-source model, asking, “Is this another DeepSeek moment? Should we expect [one] every couple of months now?”

Technical capabilities and limitations

Moonshot AI researchers said Kimi K2 Thinking set “new records across benchmarks that assess reasoning, coding and agent capabilities”. The model can execute up to 200-300 sequential tool calls without human interference, reasoning coherently across hundreds of steps to solve complex problems.

Independent testing by consultancy Artificial Analysis placed Kimi K2 on top of its Tau-2 Bench Telecom agentic benchmark with 93% accuracy, which was described as the highest score it has independently measured.

However, Nathan Lambert, a researcher at the Allen Institute for AI, suggested there’s still a time lag of approximately four to six months in raw performance between the best closed and open models, though he acknowledged that Chinese labs are closing in and performing very strongly on key benchmarks.

Market implications and competitive pressure

Zhang Ruiwang, a Beijing-based information technology system architect, said the trend was for Chinese companies to keep costs down, explaining, “The overall performance of Chinese models still lags behind top US models, so they have to compete in the realms of cost-effectiveness to have a way out”.

Zhang Yi, chief analyst at consultancy iiMedia, said the training costs of Chinese AI models were seeing a “cliff-like drop” driven by innovation in model architecture and training technique, and input of quality training data, marking a shift away from the heaping of computing resources in the early days.

The model was released under a Modified MIT License that grants full commercial and derivative rights, with one restriction: deployers serving over 100 million monthly active users or generating over US$20 million per month in revenue must prominently display “Kimi K2” on the product’s user interface.

Industry response and future outlook

Deedy Das, a partner at early-stage venture capital firm Menlo Ventures, wrote in a post on X that “Today is a turning point in AI. A Chinese open-source model is #1. Seminal moment in AI”.

🚨 Today is a turning point in AI. A Chinese open source model is #1.

Kimi K2 Thinking scored 51% in Humanity’s Last Exam, higher than GPT-5 and every other model. $0.6/M in, $2.5/M output.

The best at writing, and does 15tps on two Mac M3 Ultras!

Seminal moment in AI.

Try it… pic.twitter.com/fmxlxpCGbE

— Deedy (@deedydas) November 7, 2025

Nathan Lambert wrote in a Substack article that the success of Chinese open-source AI developers, including Moonshot AI and DeepSeek, showed how they “made the closed labs sweat,” adding “There’s serious pricing pressure and expectations that [the US developers] need to manage”.

The release positions Moonshot AI alongside other Chinese AI companies like DeepSeek, Qwen, and Baichuan that are increasingly challenging the narrative of American AI supremacy through cost-efficient innovation and open-source development strategies. 

Whether this represents a sustainable competitive advantage or a temporary convergence in capabilities remains to be seen as both US and Chinese companies continue advancing their models.

the public nature of the statements, and the market’s reaction, suggest substantive discussions may soon be underway.

The AI chip landscape is entering a period of flux. Organisations should maintain flexibility in their infrastructure strategy and monitor how partnerships like Tesla-Intel might reshape the competitive dynamics of AI hardware manufacturing.

The decisions made today about chip manufacturing partnerships could determine which organisations have access to cost-effective, high-performance AI infrastructure in the coming years.

Photo by Moonshot AI)

See also: DeepSeek disruption: Chinese AI innovation narrows global technology divide

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.

New Post has been published on https://thedigitalinsider.com/the-sequence-radar-751-last-week-in-ai-k2s-brains-lambdas-capacity-arr-gravitas/

The Sequence Radar #751: Last Week in AI: K2’s Brains, Lambda’s Capacity, ARR Gravitas

An amazing new model that pushes the boundaries of reasoning plus more deals and ARR news.

Created Using GPT-5

Next Week in The Sequence:

We continue our series about synthetic data generation with an exploration of the different types of synthetic data. Our AI of the week covers the amazing Kimi K2 Thinking. The opinion section explores the state of memory in foundation models.

Subscribe and don’t miss out:

TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

📝 Editorial: Last Week in AI: K2’s Brains, Lambda’s Capacity, ARR Gravit

This week crystallized three threads—technical progress, compute access, and business scale—that define where AI is heading.

Moonshot’s Kimi K2 is the week’s purest technical release. It pushes the Mixture-of-Experts playbook further: a huge total parameter budget, a relatively small number of experts “activated” per token, and a training stack centered on stability and agentic post-training. The “K2 Thinking” variant leans hard into long-horizon reasoning and tool use, with credible jumps on coding and logic tasks. Two takeaways matter. First, open weights plus a transparent training recipe give the community something to study and remix instead of guessing at secret sauce. Second, K2 is a concrete demonstration that MoE architectures—paired with disciplined data and post-training—can match or beat dense giants on cost/perf without relying on brute force. If you care about reproducibility and TCO, this is a north star, not a novelty.

On the industrial side, Lambda’s large-compute pact with Microsoft is the clearest sign that GPU scarcity is being professionalized into bookable, multi-year capacity. The shape of the deal is straightforward: a diversified pipeline of top-tier accelerators, delivered under a contract that smooths supply shocks and shortens time-to-capacity for model teams. Translation for practitioners: fewer roulette spins to secure training windows, more predictable paths for scaled fine-tuning and serving, and the beginning of a healthier spot market for bursts. Translation for startups: access to frontier-class clusters is drifting from “who you know” toward “what you can reserve,” which levels the playing field—at least a little—against hyperscaler lock-in.

Then there’s the money talk. OpenAI and Anthropic both projected eye-popping revenue run-rates, reframing AI as a utility build-out more than a software SKU. The headline isn’t just the numbers; it’s the operating model behind them. Premium agentic capabilities—code assistants, retrieval-augmented systems, and tool-driven workflows—are converting into durable enterprise spend. That, in turn, justifies long-cycle bets on data centers, power, and supply chains that would make a telecom CFO nod in recognition. It also resets expectations for everyone else: if the leaders are locking in capacity and turning usage into recurring revenue, the strategic gap isn’t only model quality; it’s infrastructure, go-to-market, and governance.

That’s the week: one open model that matters, one supply deal that changes access, and two revenue signals that justify the capex—and remind us that the center of gravity is shifting toward teams who can turn reasoning into reliable, governed workflows at scale.

🔎 AI Research

Scaling Agent Learning via Experience Synthesis

Authors: Meta Superintelligence Labs; FAIR at Meta; University of Chicago; UC Berkeley.

Summary: The paper introduces DreamGym, a unified RL framework that replaces costly real-environment rollouts with a reasoning-based experience model, an active replay buffer, and a curriculum task generator to synthesize diverse, causally grounded trajectories for agent training. Across WebShop, ALFWorld, and WebArena, DreamGym matches strong RL baselines using only synthetic interactions and delivers sizable sim-to-real gains while requiring far fewer real-world rollouts.

Nested Learning: The Illusion of Deep Learning Architectures

Authors: Google Research (USA).

Summary: The paper proposes Nested Learning (NL), a paradigm that treats modern models—including optimizers—as systems of nested, multi-level optimization problems with their own “context flows,” explaining in-context learning as compression of context and suggesting added “levels” for higher-order abilities. Using this lens, the authors introduce richer “deep” optimizers, a self-modifying sequence model, and a continuum memory system that together form the HOPE module, which shows strong results on language modeling and commonsense reasoning benchmarks.

CodeClash: Benchmarking Goal-Oriented Software Engineering

Authors: Stanford University; Princeton University; Cornell University.

Summary: The paper introduces CodeClash, a tournament-style benchmark where LMs iteratively edit codebases that then compete head-to-head in arenas (e.g., BattleSnake, Poker, RoboCode) to optimize high-level objectives—revealing capabilities beyond unit-test correctness. Across 1,680 tournaments, models showed creativity but shared failures in strategic reasoning and codebase maintenance; notably, top models lost every round to an expert human bot.

LiveTradeBench: Seeking Real-World Alpha with Large Language Models

Authors: University of Illinois Urbana–Champaign.

Summary: LiveTradeBench is a live, multi-market trading environment (U.S. equities and Polymarket) that streams prices/news and evaluates LLM agents on portfolio-allocation decisions, exposing gaps between static benchmark scores and real-world decision-making. Over 50 live trading days with 21 LLMs, performance in one market didn’t generalize to another and high LMArena scores didn’t predict superior trading outcomes.

RedCodeAgent: Automatic Red-Teaming Agent Against Diverse Code Agents

Authors: University of Chicago; University of Illinois Urbana–Champaign; VirtueAI; Microsoft Research; UK AI Safety Institute; University of Oxford; UC Berkeley.

Summary: RedCodeAgent is an automated red-teaming system that learns from past attacks (memory), combines multiple jailbreak tools (including code-substitution), and uses sandboxed execution to discover vulnerabilities in code agents beyond static benchmarks. It consistently outperforms baseline jailbreak methods across many risky scenarios and languages, while remaining efficient and uncovering new vulnerabilities in real-world assistants like Cursor and Codeium.

Towards a Future Space-Based, Highly Scalable AI Infrastructure System Design (Project Suncatcher)

Authors: Google Research.

Summary: Google proposes solar-powered “data centers” in space—constellations of satellites with free-space optical links and radiation-tested TPUs—to tap near-continuous solar energy and reduce terrestrial resource strain; formation-flying and short-range DWDM optical links are key enablers. Early analyses show feasibility across inter-satellite bandwidth, orbital control, TPU radiation tolerance, and launch-cost trajectories that could drop to ≲$200/kg to LEO by the mid-2030s.

🤖 AI Tech Releases

Kimi K2 Thinking

Moonshot AI released Kimi K2 Thinking, a reasoning model that excels in agentic tasks.

Magentic Marketplace

Microsoft open sourced Magentic Marketplace, an open source simulation environment for agentic markets.

📡AI Radar

OpenAI CEO Sam Altman says the company expects to end 2025 above $20B ARR and has roughly $1.4T in data-center commitments over the next eight years (statement on X).

Amazon unveils Kindle Translate (beta) to help KDP authors publish AI-translated ebooks (English↔Spanish; German→English to start).

Inception raises $50M to build diffusion-based LLMs for code and text, aiming for big latency/efficiency gains.

Snap and Perplexity sign a partnership to bring conversational AI search into Snapchat.

Replika founder Eugenia Kuyda launches Wabi (“YouTube for apps”) with a $20M pre-seed.

SoftBank and OpenAI form “SB OAI Japan,” a joint venture to market the Crystal/“Cristal” enterprise AI offering in Japan starting 2026.

Anthropic’s internal plan (as reported) targets up to $70B revenue and $17B cash flow in 2028, driven by B2B demand.

Lambda announces a multibillion-dollar, multi-year agreement with Microsoft to deploy AI infrastructure using tens of thousands of NVIDIA GPUs.

Poolside: following its Project Horizon announcement (a 2-GW Texas AI campus with CoreWeave as anchor), reports say NVIDIA is weighing up to a $1B investment. (Company context + report.) (poolside.ai)

New Post has been published on https://thedigitalinsider.com/10-of-nvidias-cost-why-tesla-intel-chip-partnership-demands-attention/

10% of Nvidia's cost: Why Tesla-Intel chip partnership demands attention

The potential Tesla-Intel chip partnership could deliver AI chips at just 10% of Nvidia’s cost – a claim that represents a significant development in AI infrastructure that enterprise technology leaders cannot afford to ignore.

On November 6, 2025, Tesla CEO Elon Musk stated publicly at the company’s annual shareholder meeting that the electric vehicle manufacturer is considering working with Intel to produce its fifth-generation AI chips, signalling a major strategic shift in how AI computing hardware might be manufactured and distributed.

“You know, maybe we’ll, we’ll do something with Intel,” Musk told shareholders, according to a Reuters report. “We haven’t signed any deal, but it’s probably worth having discussions with Intel.” The statement sent Intel shares up 4% in after-hours trading, underscoring how seriously the market views the potential collaboration.

The strategic context behind the partnership

Tesla’s consideration of Intel as a manufacturing partner comes at a important juncture for both companies. Tesla is designing its AI5 chip to power its autonomous driving systems.

Currently on its fourth-generation chip, Tesla has identified a significant supply constraint that traditional partnerships with Taiwan’s TSMC and South Korea’s Samsung cannot address fully.

“Even when we extrapolate the best-case scenario for chip production from our suppliers, it’s still not enough,” Musk said during the shareholder meeting. The supply gap has led Tesla to consider building what Musk calls a “terafab” – a massive chip fabrication facility capable of producing at least 100,000 wafer starts per month.

For Intel, the potential partnership offers an important opportunity. The US chipmaker has lagged significantly behind Nvidia in the AI chip race and desperately needs external customers for its newest manufacturing technology.

The US government recently took a 10% stake in Intel, underscoring the strategic importance of maintaining domestic chip manufacturing capabilities.

Cost and performance implications

At 10% of Nvidia’s manufacturing cost, the technical specifications Musk outlined during the shareholder meeting could reshape enterprise AI economics. According to Musk, Tesla’s AI5 chip would consume approximately one-third of the power used by Nvidia’s flagship Blackwell chip, and cost just 10% as much to manufacture.

“I’m super hardcore on chips right now, as you may be able to tell,” Musk said. “I have chips on the brain.”

The cost and efficiency projections, if realised, could alter the economics of AI deployment. Enterprise leaders investing heavily in AI infrastructure should monitor whether these performance targets materialise, as they could influence future technology purchasing decisions in the industry.

The chip would be inexpensive, power-efficient, and optimised for Tesla’s own software, Musk said.

Production timeline and scale

Tesla’s chip production roadmap provides a timeline for enterprise planning. A small number of AI5 units would be produced in 2026, with high-volume production possible in 2027. Musk indicated in a post on social media that AI6 will use the same fabrication facilities but achieve roughly twice the performance, with volume production targeted for mid-2028.

The scale of Tesla’s ambitions is substantial. The proposed “terafab” would represent an expansion of domestic chip manufacturing capacity, potentially reducing supply chain vulnerabilities that have plagued the technology industry in recent years.

“So I think we may have to do a Tesla terafab. It’s like a giga but way bigger. I can’t see any other way to get to the volume of chips that we’re looking for. So I think we’re probably going to have to build a gigantic chip fab. It’s got to be done,” Musk said.

What this means for enterprise decision-makers

Several strategic considerations emerge from any potential Tesla-Intel chip partnership:

Supply chain resilience: The move toward domestic chip manufacturing addresses concerns about supply chain concentration in Asia. Enterprise leaders managing technology risk should consider how shifts in chip manufacturing geography might affect their supply chains and vendor relationships.

Cost structure changes: If Tesla achieves its stated cost targets, the competitive landscape for AI chips could shift. Organisations should prepare contingency plans for potential price pressure on current suppliers and evaluate whether alternative chip architectures are viable.

Technology sovereignty: The US government’s stake in Intel and support for domestic chip manufacturing reflect broader geopolitical considerations. Enterprise leaders in regulated industries or those handling sensitive data should assess how the trends might affect their technology sources.

Innovation pace: Tesla’s aggressive timeline for multiple chip generations suggests an accelerating pace of AI hardware innovation. Technology leaders should factor this into refresh cycles and architecture decisions, avoiding premature commitment to current-generation technology.

The broader industry context

Musk’s statements occur against the backdrop of US-China technology competition. Export restrictions have impacted Nvidia’s business in China, where its market share has reportedly dropped from 95% to near zero.

Intel declined to comment on Musk’s remarks, and no formal agreement has been announced. However, the public nature of the statements, and the market’s reaction, suggest substantive discussions may soon be underway.

The AI chip landscape is entering a period of flux. Organisations should maintain flexibility in their infrastructure strategy and monitor how partnerships like Tesla-Intel might reshape the competitive dynamics of AI hardware manufacturing.

The decisions made today about chip manufacturing partnerships could determine which organisations have access to cost-effective, high-performance AI infrastructure in the coming years.

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.

New Post has been published on https://thedigitalinsider.com/upgrade-your-productions-with-tricaster-bundles/

Upgrade Your Productions with TriCaster Bundles

November 10, 2025

Share

Share

NDI November is a month-long, free, virtual event series dedicated to exploring the power of NDI technology for modern live production. Throughout November 2025, we’ll bring together industry experts, vendors, dealers, and system integrators to deliver an in-depth look at NDI workflows and innovations shaping the future of IP-based production. Join Vizrt on November 11th at 3pm EST to see how Vizrt TriCasters can fit into your workflow Register for FREE now and don’t miss this ultimate NDI event!

TriCaster Mini X gives producers at any level the freedom to create and share video wherever and whenever they want – truly demonstrating the power of software-defined visual storytelling. The ideal traveling partner for TriCaster Mini X, the Vizrt Control Surface, as a bundle provides studio-style control and a small footprint to deliver professional results…from the office, an event, or anywhere. Call Videoguys at 800-323-2325 for free tech advice!

Save $600!

The compact, plug-and-play design, intuitive layout mapped to the TriCaster Mini X interface, large backlit buttons, and premium T-Bar make this controller a perfect match for faster, more precise, and comfortable live production.

$10,595.00 reg.

$9,995.00 PROMO

Offer expires 12/31/25

Accessibility, Unlocked Bringing the possibilities of professional live video production to anyone with a story to tell – without vast investments in infrastructure. Flexibility, Attained Off-the-shelf HDMI devices connect directly to the Mini X in minutes creating professional-level productions without having to purchase any new equipment. Scalability, Achieved As part of the Vizrt ecosystem of products, users can take advantage of the many ways to scale up their productions to suit any need.

Save $200!

Easy to set-up and with hundreds of capabilities, TriCaster Mini X puts expert tools at your fingertips, without you having to be a production expert.

$8,195.00 reg.

$7,995.00 PROMO

Offer expires 12/31/25

Take your production to the next level with TriCaster and Flex—the winning combination for professional live video workflows. Choose from five powerful configurations tailored to fit any production environment and save big when you bundle. Call Videoguys at 800-323-2325 for free expert advice, help finding the perfect setup, or to connect with a local integrator from our nationwide dealer network. Offer expires Dec 31st, 2025

A compact solution built for small to mid-sized productions with essential IP workflows.

$19,995.00 reg.

$15,495.00 PROMO

Offer expires 12/31/25

A scalable system designed for larger productions, offering 16 external inputs and UHD switching.

$25,490.00 reg.

$20,995.00 PROMO

Offer expires 12/31/25

An advanced platform with premium features like Live Call Connect, Live Story Creator, and real-time remote workflows.

$32,490.00 reg.

$27,995.00 PROMO

Offer expires 12/31/25

A rackmount system tailored for professional studios, supporting 8-channel switching with powerful automation.

$45,490.00 reg.

$30,995.00 PROMO

Offer expires 12/31/25

A rackmount system tailored for professional studios, supporting 16-channel switching with powerful automation.

$45,490.00 reg.

$32,995.00 PROMO

Offer expires 12/31/25

A high-performance tower system built for enterprise-level productions, delivering 16-channel switching, UHD capabilities, and maximum expandability.

$47,490.00 reg.

$32,995.00 PROMO

Offer expires 12/31/25

New Post has been published on https://thedigitalinsider.com/headings-semantics-fluidity-and-styling-oh-my/

Headings: Semantics, Fluidity, and Styling — Oh My!

A few links about headings that I’ve had stored under my top hat.

“Page headings don’t belong in the header”

Martin Underhill:

I’ll start with where the <h1> should be placed, and you’ll start to see why the <header> isn’t the right location: it’s the header for the page, and the main page content should live within the <main> element.

A classic conundrum! I’ve seen the main page heading (<h1>) placed in all kinds of places, such as:

The site <header> (wrapping the site title)

A <header> nested in the <main> content

A dedicated <header> outside the <main> content

Aside from that first one — the site title serves a different purpose than the page title — Martin pokes at the other two structures, describing how the implicit semantics impact the usability of assistive tech, like screen readers. A <header> is a wrapper for introductory content that may contain a heading element (in addition to other types of elements). Similarly, a heading might be considered part of the <main> content rather than its own entity.

So:

<!-- 1️⃣ --> <header> <!-- Header stuff --> <h1>Page heading</h1> </header> <main> <!-- Main page content --> </main> <!-- 2️⃣ --> <main> <header> <!-- Header stuff --> <h1>Page heading</h1> </header> <!-- Main page content --> </main>

Like many of the decisions we make in our work, there are implications:

If the heading is in a <header> that is outside of the <main> element, it’s possible that a user will completely miss the heading if they jump to the main content using a skip link. Or, a screenreader user might miss it when navigating by landmark. Of course, it’s possible that there’s no harm done if the first user sees the heading prior to skipping, or if the screenreader user is given the page <title> prior to jumping landmarks. But, at worst, the screenreader will announce additional information about reaching the end of the banner (<header> maps to role="banner") before getting to the main content.

If the heading is in a <header> that is nested inside the <main> element, the <header> loses its semantics, effectively becoming a generic <div> or <section>, thus introducing confusion as far as where the main page header landmark is when using a screenreader.

All of which leads to Martin to a third approach, where the heading should be directly in the <main> content, outside of the <header>:

<!-- 3️⃣ --> <header> <!-- Header stuff --> </header> <main> <h1>Page heading</h1> <!-- Main page content --> </main>

This way:

The <header> landmark is preserved (as well as its role).

The <h1> is connected to the <main> content.

Navigating between the <header> and <main> is predictable and consistent.

As Martin notes: “I’m really nit-picking here, but it’s important to think about things beyond the visually obvious.”

“Fluid Headings”

Donnie D’Amato:

There’s no shortage of posts that explain how to perform responsive typography. […] However, in those articles no one really mentions what qualities you are meant to look out for when figuring out the values. […] The recommendation there is to always include a non-viewport unit in the calculation with your viewport unit.

To recap, we’re talking about text that scales with the viewport size. That usually done with the clamp() function, which sets an “ideal” font size that’s locked between a minimum value and a maximum value it can’t exceed.

.article-heading font-size: clamp(<min>, <ideal>, <max>);

As Donnie explains, it’s common to base the minimum and maximum values on actual font sizing:

.article-heading font-size: clamp(18px, <ideal>, 36px);

…and the middle “ideal” value in viewport units for fluidity between the min and max values:

.article-heading font-size: clamp(18px, 4vw, 36px);

But the issue here, as explained by Maxwell Barvian on Smashing Magazine, is that this muffs up accessibility if the user applies zooming on the page. Maxwell’s idea is to use a non-viewport unit for the middle “ideal” value so that the font size scales to the user’s settings.

Donnie’s idea is to calculate the middle value as the difference between the min and max values and make it relative to the difference between the maximum number of characters per line (something between 40-80 characters) and the smallest viewport size you want to support (likely 320px which is what we traditionally associate with smaller mobile devices), converted to rem units, which .

.article-heading --heading-smallest: 2.5rem; --heading-largest: 5rem; --m: calc( (var(--heading-largest) - var(--heading-smallest)) / (30 - 20) /* 30rem - 20rem */ ); font-size: clamp( var(--heading-smallest), var(--m) * 100vw, var(--heading-largest) );

I couldn’t get this working. It did work when swapping in the unit-less values with rem. But Chrome and Safari only. Firefox must not like dividing units by other units… which makes sense because that matches what’s in the spec.

Anyway, here’s how that looks when it works, at least in Chrome and Safari.

Style :headings

Speaking of Firefox, here’s something that recently landed in Nightly, but nowhere else just yet.

Alvaro Montoro:

Styling headings in CSS is about to get much easier. With the new :heading pseudo-class and :heading() function, you can target headings in a cleaner and more flexible way.

:heading: Selects all <h*> elements.

:heading(): Same deal, but can select certain headings instead of all.

I scratched my head wondering why we’d need either of these. Alvaro says right in the intro they select headings in a cleaner, more flexible way. So, sure, this:

:heading

…is much cleaner than this:

h1, h2, h3, h4, h5, h6

Just as:

:heading(2, 3)

…is a little cleaner (but no shorter) than this:

h2, h3

New Post has been published on https://thedigitalinsider.com/the-universal-tool-calling-protocol-for-agentic-ai/

The universal tool calling protocol for agentic AI

How many of you have actually built something with tool calling? If you’re reading this, chances are you’ve at least heard about it. Maybe you’ve integrated a few tools into your AI agents, or perhaps you’re on the other side, providing tools for agents to use. Either way, you know that tool calling is what separates a chatbot from a true AI agent.

Here’s the thing: agents without tools are just fancy text generators. What makes them genuinely useful, what gives them their power, is their ability to reach beyond the confines of language models and actually do things in the real world. They can read your emails, browse the web, manipulate files, and interact with the countless APIs that power our digital infrastructure.

But we have a problem. A big one.

The integration bottleneck we all know too well

Picture this: you’re building an AI agent, and you want it to interact with Gmail, browse the web, and work with files on a user’s computer. In the early days, you’d have to code each integration yourself. Every. Single. One.

This approach created an impossible situation. Agent providers became bottlenecks in their own ecosystems. Want to add a new tool? Sorry, you’ll have to wait for the provider to build that integration. Have a proprietary API that’s specific to your business? Good luck getting that on anyone’s roadmap.

The community recognized this problem and came up with a solution: the Model Context Protocol (MCP). The promise was elegant: standardize how agents communicate with tools so that agent providers only need to implement one communication protocol. Then, any tool can provide a server that translates its functionality into a model-friendly format.

Sounds great, right? Well, not so fast.

This post is for paying subscribers only

Subscribe now

Already have an account? Sign in

Become a member to see the rest.

You’ve landed on a piece of content that’s exclusive to members. Members have access to templates, real-world presentations, events, reports, salary calculators, and more. Not yet a member? Sign up for free.

Sign up See all plans

New Post has been published on https://thedigitalinsider.com/ai-agents-5-lessons-for-getting-it-right/

AI agents: 5 lessons for getting it right

According to Gartner, over 40% of agentic AI projects will be canceled by the end of 2027. The core issue isn’t the technology itself, but how organizations implement it.

AI agents represent a new generation of automation: systems capable of completing tasks with minimal human intervention. But as companies move from pilot to production, many encounter a gap between expectations and real-world outcomes.

Based on case studies, industry examples, and lessons from practice, here are five lessons for deploying AI agents successfully.

1. Align strategy across the organization

Companies typically approach AI agents from two directions: executive mandates or isolated team experiments. Neither works well in isolation.

Bottom-up initiatives often generate promising pilots, but without executive sponsorship, they rarely scale. I’ve seen colleagues build prototypes that improved workflows and unlocked new features, but without budget or leadership support, they stalled. 

This pattern isn’t unique, Mckinsey reports fewer than 30% of companies have CEO-level sponsorship for AI, leading to scattered micro-initiatives with little enterprise impact.

Top-down efforts can also fail. For example, Salesforce CEO Marc Benioff claimed AI now performs 30–50% of company work and envisioned 1 billion agents by the end of the year. 

The statement sparked criticism. Employees argued it overstated current AI capabilities and downplayed human contributions.

The solution? Combine both approaches strategically:

Start with executives defining clear objectives and success metrics. Have technical teams run discovery workshops to assess feasibility. Then launch small pilots with executive sponsors who remove blockers without micromanaging. 

In my current work on building an AI agent system, we followed this blended approach. Executives set the vision and goals. Our team focused on building AI skills through structured learning and hands-on experimentation. 

We validated ideas with proofs of concept, iterated quickly, and created space for ongoing learning on both sides. 

When skepticism arose, we addressed concerns directly before moving forward. The AI Agents space is moving fast, and the pace of mutual adaptation is critical to maintain momentum.

The key is creating a bridge between strategic vision and operational reality. Something that requires both top-down support and bottom-up expertise.

CV algorithm development by the masses for the masses

Discover how CV algorithm development is becoming more accessible, enabling more people to build & benefit from real-world CV solutions at scale.

2. Address data readiness early

AI agents don’t generate new knowledge, they operate on available information. For most organizations, that information is fragmented and unstructured.

This is the rule of thumb I use when assessing readiness: if you can’t access 80% of relevant data programmatically, or more than 30% of critical knowledge still lives in people’s heads, you’re not ready. 

In my project building threat intelligence agents, this principle proved true, most of the effort went into consolidating data, not agent design.

The risk of poor data readiness? Your agent will hallucinate or require constant human intervention. 

Air Canada’s AI chatbot told customers about a refund policy that didn’t exist, leading to the airline being ordered to compensate passengers who received incorrect information. The tribunal ruled that Air Canada is responsible for all information on its website, including chatbot responses.

Start by capturing institutional knowledge. Then structure your unstructured data incrementally, focusing on the most critical information first. Build feedback loops to capture and fix data gaps as you discover them.

3. Set realistic performance expectations

Organizations tolerate 5-10% human error rates yet demand perfection from AI agents. This mismatch kills promising initiatives.

Media hype often sets organizations up for disappointment. I’ve seen teams hold agents to impossible standards because vendors claim ‘100% automation.’ 

For example, Accenture promotes agents that can read every insurance submission, a huge leap from today’s reality where half are still untouched. In practice, these claims raise expectations far beyond what teams can reliably deliver.

The mindset shift needed? Benchmark against human performance, not perfection. Klarna’s customer service AI demonstrates this approach: it resolves 66% of requests, reduces resolution time from 11 minutes to 2, and maintains satisfaction scores comparable to human agents. They didn’t aim for 100% – they aimed for better than human.

In developing AI systems, I’ve learned that accuracy isn’t the only metric. We start with the customer baseline and measure how much faster AI delivers value. 

Accuracy still matters, but when issues arise, we address them openly and improve incrementally. Phased rollouts; alpha, beta, and early customer feedback, help us refine performance and build trust without over-promising.

Focus on handling 80% of cases well rather than 100% perfectly. 

Start with low-risk use cases where mistakes have minimal impact. Internal knowledge searches or data validation build confidence without risking customer relationships. 

Communicate that agents provide confidence scores, not certainties. When stakeholders understand the logic, they’re more comfortable with occasional errors.

Why AI startups should bet big on privacy

Smart AI startups are turning privacy from a roadblock into their biggest competitive advantage. Here’s how they’re doing it.

4. Balance build vs. buy strategically

The build-versus-buy decision isn’t binary, it’s about finding the right hybrid approach.

Fully in-house development seems appealing but often fails. I’ve watched several organizations attempt to build agent platforms entirely in-house, only to hit walls in orchestration, memory, and governance. The reality is, few teams have the specialized expertise required, Forrester echoes this, predicting that three-quarters of such efforts will fail.

Complete outsourcing has its own pitfalls. If everyone uses the same vendor’s agent, you lose competitive advantage. Vendors optimize for common cases, not your specific needs.

The hybrid sweet spot: Start with commercial solutions to validate value quickly. Then identify what makes your use case unique, these become candidates for custom development.

Critical skills to develop internally include prompt engineering, data pipeline development, and domain expertise deep enough to guide the agent effectively. 

You don’t need to build everything, but you need to understand and control what makes you different.

5. Don’t overlook operational infrastructure

Many pilots succeed in controlled environments but fail in production – not because of the agent itself, but because of missing operational infrastructure. 

I’ve seen agents run flawlessly in notebooks and staging, only to collapse in production when a data format changed silently. Without monitoring, failures went unnoticed until real damage was done. 

Replit’s recent incident illustrates the risk: their coding agent deleted a production database despite safeguards, showing how fragile operations can be without rigorous controls.

“Building with production in mind” means considering operational requirements from day one. Before any prototype, ask: How will we know if it’s working? What happens when it fails? Who can override its decisions? How do we audit its actions?

Essential infrastructure components include:

Access controls: Who can invoke the agent and what can it modify?

Observability: Logging, metrics, and anomaly detection

Cost management: Token tracking, API quotas, and automatic shutoffs

Integration safeguards: Rate limiting, circuit breakers, and graceful degradation

Incident response: Kill switches, rollback procedures, and escalation paths

Start testing in notebooks, then staging with synthetic data, then shadow mode alongside human processes, before limited production rollout. Each stage reveals different challenges and builds operational confidence.

Final thoughts: Move early, learn continuously

Successful AI agent adoption isn’t about perfect technology or massive budgets. It’s about organizational learning speed. 

Companies that succeed start before everything is perfect, build incrementally, fail fast, and scale what works.

New Post has been published on https://thedigitalinsider.com/why-agentic-ai-is-the-future-of-virtual-assistants/

Why agentic AI is the future of virtual assistants

You know that feeling when you call customer support and the agent just… doesn’t get it? They’re reading from a script, asking you to repeat steps you’ve already tried, completely missing the frustration in your voice.

Now imagine if that agent could actually see you’re upset, understand what you’re trying to achieve, and adapt their approach accordingly. That’s the gap between today’s automated systems and what virtual assistants should actually be.

I’m Raj, and I’ve spent my entire professional life researching how we learn from what we see, hear, and observe.

Today, I want to share what I’ve learned about building virtual assistants that actually work, not just automated processes that frustrate users, but genuine collaborative partners that understand context, show empathy, and build trust.

The problem with today’s “agents”

Let’s be honest: most of what we call AI agents today are just glorified robotic processes. We had those before AI became the buzzword du jour. They follow predetermined paths, match patterns to intents, and spit out pre-programmed responses. But is that really what we need?

Think about real-life agents, the human ones. Whether you’re talking to a customer support representative, a healthcare professional, or a financial advisor, there’s actual collaboration happening. They understand not just what you’re saying, but why you’re saying it. They pick up on your mood, adapt their approach, and work with you toward your goals.

The missing piece? Theory of mind.

For those unfamiliar with the concept, the theory of mind is our ability to understand that others have beliefs, desires, and intentions different from our own.

When someone talks to you, you’re not just processing their words; you’re assessing their goals, understanding their beliefs, and figuring out how to help them based on what you know to be true. It’s not about pattern recognition or intent mapping. It’s about genuine understanding.

The four pillars of effective virtual assistants

Through our work developing EVA (our Enterprise Virtual Assistant), we’ve identified four essential phases that any effective virtual assistant must master:

1. Knowledge acquisition: More than Just RAG

First things first: to help anyone with anything, you need knowledge. But here’s the thing: acquiring and utilizing enterprise knowledge remains a massive challenge. Sure, we have structured databases, unstructured documents, and various repositories of information.

But RAG (Retrieval-Augmented Generation)? It’s really just a glorified search mechanism.

Real knowledge acquisition means understanding predicates, actions, and applicable conditions that aren’t explicitly written anywhere. Take credit card fraud, for example. You need to report it within 24 hours for the bank to waive charges. But that information might be buried in legal documents, and the system needs to understand when to surface it based on context.

2. Conversation: Beyond information retrieval

When you ask a virtual assistant a question, are you just looking for information retrieval? Usually not. You want a conversation; a back-and-forth that helps you solve a problem or achieve a goal.

Let me give you my favorite example: “If my top five customers’ sentiment falls below 5%, schedule a call with my northeast sales team.”

Sounds simple? It’s not. The system needs to understand:

What customer sentiment means and where to find it

How to calculate a 5% drop

That “northeast” is a geographical region

Which team members are assigned to that region

How to access scheduling systems

This isn’t scripting; it’s understanding context and taking appropriate action.

3. Agency: Multi-step problem solving

Real agency means handling complex, multi-step tasks without explicit programming for each scenario. When someone says, “I hit a wall with my car,” why do you think they’re calling their insurance company? Obviously, they want to file a claim and remedy the situation.

A truly intelligent agent recognizes the negative state and navigates the user to a positive outcome. Like a GPS recalculating when you miss an exit, it adapts dynamically based on your current situation and ultimate goal. It doesn’t say, “I told you to follow my instructions.” It simply recalculates and guides you forward.

4. Empathy and trust: The human touch

Here’s what everyone seems to forget: AI use cases will be severely limited without empathy and trust. Trust comes from reasoning and providing certified, factual information. Empathy comes from understanding and responding appropriately to emotional context.

Imagine a florist’s virtual assistant. When someone mentions they need flowers for their daughter’s graduation, the response should be jubilant and celebratory. But if they’re ordering for a funeral? The entire tone needs to shift to something more somber and respectful.

Nobody wants to talk to a mechanical-sounding agent with no emotional intelligence. I’m not saying we need to anthropomorphize these systems into virtual girlfriends or boyfriends, but they do need to engage at a human level.

AI agents: 5 lessons for getting it right

Based on case studies, industry examples, and lessons from practice, here are five lessons for deploying AI agents successfully.

The architecture of understanding

So how do we build systems that can actually do all this? The answer lies in what we call neurosymbolic systems: combining the scale of deep learning with the reliability of symbolic reasoning.

Look, I know there’s debate about this. Some folks think transformer models and deep learning will eventually handle everything. But right now, for complex cognitive tasks, pure deep learning just isn’t cutting it.

My daughter figured this out after one day of playing with large language models. She noticed they repeat stories, creating sentences that sound coherent but often lack real meaning.

Neurosymbolic systems give us:

Scale from deep learning approaches

Reliability from symbolic reasoning

Explainability for trust-building

Factual grounding to prevent hallucination

When you extract information into graph representations with known relationships, traversing that graph is like querying a database – you know the information is true. No hallucination, no made-up facts.

Multimodal understanding: Seeing beyond words

Here’s where things get really interesting. Real communication is about everything else, too. When I’m giving a presentation and see everyone checking their phones, should I just keep talking? Of course not. That visual feedback tells me I need to change my approach.

Our virtual assistants need the same awareness. They should know:

Whether someone is present in their field of view

If the user is engaged or distracted

Environmental factors (like being on mute during a call)

Emotional states through facial expressions

Even personality traits that emerge over time

We’ve built systems that can assess mental health conditions with 85% accuracy compared to human experts in just five minutes. How? By analyzing not just what people say, but how they say it.

When you’re recalling difficult memories, emotions express themselves in facial micro-expressions that you can’t conceal. Your spouse can read these signals, so why shouldn’t your virtual assistant?

Real-world applications today

This isn’t just theoretical. We have customers using multimodal virtual assistants for:

Damage assessment after storms

Safety inspections in restaurants and facilities

Vehicle inspection verification

Mental health screening for deployment readiness

Real-time compliance monitoring

These systems combine enterprise knowledge with real-world observation. They understand regulations, observe actual conditions, and assess violations or compliance in real-time.

For instance, detecting a person, a phone, and a car isn’t the point. Understanding that someone is driving while talking on the phone – that’s what constitutes a violation. The system needs to understand relationships, not just identify objects.

The challenge of exponential information growth

Here’s something that should keep you up at night: data is doubling every twelve hours. Let that sink in. Without AI assistance, we’ll actively look dumber as we fall further behind the information curve.

But here’s the kicker: much of this “new” data isn’t original content. AI agents are competing to generate synthetic content, muddying the waters further. Model drift is coming, and it’s going to be a serious problem.

That’s why, at least for the near term, we need neurosymbolic systems grounded in truth. Systems that can:

Process information multimodally

Engage with genuine empathy

Build and maintain trust

Deliver measurable ROI through better engagement

Why AI startups should bet big on privacy

Smart AI startups are turning privacy from a roadblock into their biggest competitive advantage. Here’s how they’re doing it.

Building for the future

Six months from now, you’ll see the rebirth of wearable technology; not just watches, but glasses and other immersive devices. People will walk through the world asking questions and getting real-time assistance. Privacy concerns aside (and yes, that’s a whole other conversation), these devices will fundamentally change how we interact with AI.

Imagine walking through a construction site with smart glasses, getting real-time safety assessments. Or a doctor examining a patient while an AI assistant observes symptoms and suggests diagnostic paths based on visual and verbal cues.

The path forward

The virtual assistants of tomorrow will truly assist. They’ll understand context, show appropriate emotion, and build trust through reliable, explainable actions. They’ll see when you’re frustrated, hear the stress in your voice, and adapt their approach accordingly.

This is about building systems that understand human communication in all its forms, verbal, visual, and emotional, and respond appropriately. It’s about moving beyond pattern matching to genuine understanding.

The technology exists. We’ve proven it works. Now it’s time to implement it at scale, creating virtual assistants that don’t just automate processes but genuinely collaborate with humans to achieve better outcomes.

Your CFO wants ROI? Better engagement scores, higher customer satisfaction, and more efficient problem resolution – that’s the return on building virtual assistants with empathy and understanding. Your customers want to feel heard and helped? That requires systems that can see, understand, and respond with appropriate emotional intelligence.

The age of mechanical, scripted responses is ending. The era of empathetic, intelligent virtual assistants has begun. The question is about how quickly you can implement it before your competitors do.

New Post has been published on https://thedigitalinsider.com/nvidia-ai-chip-ban-can-tech-giants-navigate-a-geopolitical-zero-sum-game/

Nvidia AI chip ban: Can tech giants navigate a geopolitical zero-sum game?

When Nvidia CEO Jensen Huang initially told the Financial Times that China would “win the AI race” before softening his stance, it crystallised a predicament that’s been years in the making. The world’s most valuable chipmaker now finds itself caught between two superpowers, each wielding the Nvidia AI chip ban as a weapon in a broader technological cold war—and the company’s attempt to please both sides may ultimately satisfy neither.

From dominance to zero: A market collapse

The numbers tell a stark story. Speaking at a Citadel Securities event in October, Huang revealed that Nvidia’s share of China’s AI accelerator market has collapsed from roughly 95% to zero, with the company now assuming no revenue from China in its forecasts. This isn’t just a revenue hiccup—China previously represented between 20% and 25% of Nvidia’s data centre revenue, a segment that generated more than US$41 billion in its most recent financial results.

The latest blow came this week when sources claimed that the White House informed federal agencies it will not permit Nvidia to sell its latest scaled-down AI chips to China, specifically the B30A chip designed to train large language models. Despite Nvidia providing samples to Chinese customers and reportedly working to modify the design, the Trump administration has drawn a hard line.

But Washington’s restrictions represent only half of Nvidia’s problem. Beijing has issued guidance requiring new data centre projects receiving state funds to use only domestically-made AI chips, with projects less than 30% complete ordered to remove all installed foreign chips or cancel purchase plans. 

It’s a pincer movement that leaves Nvidia with virtually no room to manoeuvre.

The lobbying game: Too much, too late?

Huang has long argued that maintaining China’s dependence on American hardware serves US interests. His logic? Keep Chinese AI developers hooked on Nvidia’s ecosystem, and America retains technological leverage. 

Following meetings with President Trump in July, it appeared Huang’s lobbying had worked, with Washington agreeing to ease some chip curbs under a plan where Nvidia and AMD would pay the US government 15% of their Chinese revenues.

That optimism proved short-lived. Beijing has since shut Nvidia out of the market through a national security review of its chips, with Huang stating the firm’s market share has been reduced to zero. The irony is palpable: while Huang lobbied Washington to allow more sales to China, Beijing was simultaneously building barriers to keep Nvidia out.

When Huang contrasted China’s pro-industry energy subsidies with what he described as excessive Western regulation, it revealed the fundamental tension in Nvidia’s position. The company needs a favourable policy from both capitals, but operates in an environment where pleasing one increasingly means antagonising the other.

The cost of technological nationalism

This isn’t merely a corporate problem—it’s reshaping the global AI landscape. China’s ban would eliminate foreign chipmakers like Nvidia from a significant portion of the market, even if a deal is agreed to allow the resumption ofadvanced chip sales to China. 

Meanwhile, Chinese companies have over US$100 billion in state funding for AI data centre projects since 2021, creating a massive captive market for domestic alternatives.

The policy whiplash has real consequences. Following Trump’s meetings with Chinese President Xi Jinping, highly anticipated trade talks yielded no concessions from either side on chip policy, with top US officials rallying against Trump’s initial consideration of Huang’s request to allow sales of new AI chips to China.

An Nvidia spokesperson’s response to the latest restrictions was telling Reuters: “zero share in China’s highly competitive market for datacenter compute, and do not include it in our guidance”. It’s a public acknowledgement of defeat wrapped in corporate speak.

China’s calculated response

Beijing’s moves reveal a strategy that extends beyond mere retaliation. China has discouraged local tech giants from purchasing advanced Nvidia chips over security concerns this year, while showing off a new data centre powered solely by domestic AI chips. The message is clear: foreign dependence is a vulnerability to be eliminated, not managed.

The Chinese government is carving out market share for domestic chipmakers ranging from Huawei Technologies to smaller players like Shanghai-listed Cambricon and startups including MetaX, Moore Threads, and Enflame. 

While these companies struggle to match Nvidia’s performance and software ecosystem, they’re getting exactly what they need most: time, money, and a protected market to mature.

The impossible balance

Nvidia’s predicament exposes a broader truth about technology in an era of great power competition: the middle ground is disappearing. Companies can optimise for American national security priorities or Chinese market access, but increasingly not both.

Huang expressed concerns that the West was being held back by “cynicism” and excessive regulation, contrasting this with China’s energy subsidies aimed at lowering costs for local developers using domestic chips. But this comparison misses the point. 

The question isn’t whether China’s industrial policy is more effective—it’s whether Nvidia can operate in an environment where technology has become inseparable from geopolitics. The B30A saga illustrates the futility of technical compromises. 

Even a chip deliberately neutered to comply with US export controls finds no approval from Washington, while Beijing increasingly views any foreign chip as a strategic vulnerability. Nvidia could design a thousand variants, each weaker than the last, and still find itself shut out by one capital or the other.

What comes next?

In the short term, Nvidia faces a stark reality: the company now assumes 0% revenue from China in all forecasts, with Huang stating, “If anything happens in China… it will be a bonus”. This conservative guidance protects the stock but signals that management sees no near-term resolution.

The real question is whether this represents a temporary freeze or a permanent fracture. While the move helps boost sales of domestically developed chips, it also risks widening the US-China gap in AI computing power, as US tech giants continue spending hundreds of billions on data centres powered by Nvidia’s most advanced chips.

For Nvidia, the path forward likely involves doubling down on markets where geopolitics align with business—the US, Europe, and friendly Asian nations. The China dream, at least in its previous form, appears over. Huang’s softening of his “China will win” comments reflects this new reality. America might not win by keeping China dependent on its chips, but Nvidia certainly loses by being caught in the middle.

The Nvidia AI chip ban—from both directions—represents more than export controls or industrial policy. It’s evidence that in the AI race, there won’t be neutral suppliers. Technology companies will increasingly be forced to choose sides, and those who hesitate will find the choice made for them. 

Nvidia’s plunge from 95% to zero market share in China took just months. The question now is whether Washington and Beijing will leave any space for global tech companies to operate at all.

(Photo by OpenAI and Nvidia plan $100B chip deal for AI future)

See also:

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.

New Post has been published on https://thedigitalinsider.com/microsofts-next-big-ai-bet-building-a-humanist-superintelligence/

Microsoft’s next big AI bet: building a ‘humanist superintelligence’

Microsoft is forming a new team to research superintelligence and other advanced forms of artificial intelligence.

Mustafa Suleyman, who leads Microsoft’s AI division overseeing Bing and Copilot, announced the creation of the MAI Superintelligence Team in a blog post. He said he will head the group and that Microsoft plans to put “a lot of money” behind the effort.

“We are doing this to solve real, concrete problems and do it in such a way that it remains grounded and controllable,” Suleyman wrote. “We are not building an ill-defined and ethereal superintelligence; we are building a practical technology explicitly designed only to serve humanity.”

Building a ‘humanist’ approach to superintelligence

The move comes as big tech companies race to attract top AI researchers. Meta, Facebook’s parent company, recently created its own Meta Superintelligence Labs and spent billions recruiting experts, even offering signing bonuses as high as $100 million. Suleyman didn’t comment on whether Microsoft plans to match such offers but said the new team will include both internal talent and new hires, with Karen Simonyan as chief scientist.

Before joining Microsoft, Suleyman co-founded DeepMind, which Google bought in 2014. He later led the AI startup Inflection, which Microsoft acquired last year along with several of its employees.

The hiring push reflects a broader trend. Since OpenAI released ChatGPT in 2022, companies have raced to bring generative AI into their products. Microsoft uses OpenAI’s models in Bing and Copilot, while OpenAI relies on Microsoft’s Azure cloud to power its tools. Microsoft also holds a $135 billion stake in OpenAI after a recent restructuring.

Reducing reliance on OpenAI

Despite the partnership, Microsoft has been working to diversify its AI sources as it lays the groundwork for future superintelligence research. Following the Inflection acquisition, the company began experimenting with models from Google and Anthropic, another AI startup founded by former OpenAI executives.

The new Microsoft AI research group will aim to build useful AI companions that assist people in education and other areas. Suleyman said the team also plans to focus on projects in medicine and renewable energy.

A different path from rivals

Unlike some peers, Suleyman said Microsoft isn’t trying to build an “infinitely capable generalist” AI. He doubts such systems could be kept under control and instead wants to develop what he calls “humanist superintelligence” – AI that serves human needs and delivers real-world benefits.

“Humanism requires us to always ask the question: does this technology serve human interests?” he said.

While the risks of AI are widely debated – from bias to existential threats – Suleyman said his team’s goal is to create specialist systems that achieve “superhuman performance” without posing major risks. He cited examples like AI that could improve battery storage or design new molecules, similar to DeepMind’s AlphaFold project that predicts protein structures.

Medical superintelligence on the horizon

Suleyman said Microsoft is especially focused on healthcare, predicting that AI capable of expert-level diagnosis could emerge in the next two or three years.

He described it as technology that can reason through complex medical problems and detect preventable diseases much earlier. “We’ll have expert-level performance at the full range of diagnostics, alongside highly capable planning and prediction in operational clinical settings,” he wrote.

As investors question whether massive AI spending will translate into profits, Suleyman emphasised that Microsoft is setting clear limits. “We are not building a superintelligence at any cost, with no limits,” he said.

(Photo by Praswin Prakashan)

See also: Microsoft gives free Copilot AI services to US government workers

Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and co-located with other leading technology events. Click here for more information.

New Post has been published on https://thedigitalinsider.com/rl-without-td-learning/

RL without TD learning

In this post, I’ll introduce a reinforcement learning (RL) algorithm based on an “alternative” paradigm: divide and conquer. Unlike traditional methods, this algorithm is not based on temporal difference (TD) learning (which has scalability challenges), and scales well to long-horizon tasks.

We can do Reinforcement Learning (RL) based on divide and conquer, instead of temporal difference (TD) learning.

Problem setting: off-policy RL

Our problem setting is off-policy RL. Let’s briefly review what this means.

There are two classes of algorithms in RL: on-policy RL and off-policy RL. On-policy RL means we can only use fresh data collected by the current policy. In other words, we have to throw away old data each time we update the policy. Algorithms like PPO and GRPO (and policy gradient methods in general) belong to this category.

Off-policy RL means we don’t have this restriction: we can use any kind of data, including old experience, human demonstrations, Internet data, and so on. So off-policy RL is more general and flexible than on-policy RL (and of course harder!). Q-learning is the most well-known off-policy RL algorithm. In domains where data collection is expensive (e.g., robotics, dialogue systems, healthcare, etc.), we often have no choice but to use off-policy RL. That’s why it’s such an important problem.

As of 2025, I think we have reasonably good recipes for scaling up on-policy RL (e.g., PPO, GRPO, and their variants). However, we still haven’t found a “scalable” off-policy RL algorithm that scales well to complex, long-horizon tasks. Let me briefly explain why.

Two paradigms in value learning: Temporal Difference (TD) and Monte Carlo (MC)

In off-policy RL, we typically train a value function using temporal difference (TD) learning (i.e., Q-learning), with the following Bellman update rule:

[beginaligned Q(s, a) gets r + gamma max_a’ Q(s’, a’), endaligned]

The problem is this: the error in the next value $Q(s’, a’)$ propagates to the current value $Q(s, a)$ through bootstrapping, and these errors accumulate over the entire horizon. This is basically what makes TD learning struggle to scale to long-horizon tasks (see this post if you’re interested in more details).

To mitigate this problem, people have mixed TD learning with Monte Carlo (MC) returns. For example, we can do $n$-step TD learning (TD-$n$):

[beginaligned Q(s_t, a_t) gets sum_i=0^n-1 gamma^i r_t+i + gamma^n max_a’ Q(s_t+n, a’). endaligned]

Here, we use the actual Monte Carlo return (from the dataset) for the first $n$ steps, and then use the bootstrapped value for the rest of the horizon. This way, we can reduce the number of Bellman recursions by $n$ times, so errors accumulate less. In the extreme case of $n = infty$, we recover pure Monte Carlo value learning.

While this is a reasonable solution (and often works well), it is highly unsatisfactory. First, it doesn’t fundamentally solve the error accumulation problem; it only reduces the number of Bellman recursions by a constant factor ($n$). Second, as $n$ grows, we suffer from high variance and suboptimality. So we can’t just set $n$ to a large value, and need to carefully tune it for each task.

Is there a fundamentally different way to solve this problem?

The “Third” Paradigm: Divide and Conquer

My claim is that a third paradigm in value learning, divide and conquer, may provide an ideal solution to off-policy RL that scales to arbitrarily long-horizon tasks.

Divide and conquer reduces the number of Bellman recursions logarithmically.

The key idea of divide and conquer is to divide a trajectory into two equal-length segments, and combine their values to update the value of the full trajectory. This way, we can (in theory) reduce the number of Bellman recursions logarithmically (not linearly!). Moreover, it doesn’t require choosing a hyperparameter like $n$, and it doesn’t necessarily suffer from high variance or suboptimality, unlike $n$-step TD learning.

Conceptually, divide and conquer really has all the nice properties we want in value learning. So I’ve long been excited about this high-level idea. The problem was that it wasn’t clear how to actually do this in practice… until recently.

A practical algorithm

In a recent work co-led with Aditya, we made meaningful progress toward realizing and scaling up this idea. Specifically, we were able to scale up divide-and-conquer value learning to highly complex tasks (as far as I know, this is the first such work!) at least in one important class of RL problems, goal-conditioned RL. Goal-conditioned RL aims to learn a policy that can reach any state from any other state. This provides a natural divide-and-conquer structure. Let me explain this.

The structure is as follows. Let’s first assume that the dynamics is deterministic, and denote the shortest path distance (“temporal distance”) between two states $s$ and $g$ as $d^*(s, g)$. Then, it satisfies the triangle inequality:

[beginaligned d^*(s, g) leq d^*(s, w) + d^*(w, g) endaligned]

for all $s, g, w in mathcalS$.

In terms of values, we can equivalently translate this triangle inequality to the following “transitive” Bellman update rule:

[beginaligned V(s, g) gets begincases gamma^0 & textif s = g, \ gamma^1 & textif (s, g) in mathcalE, \ max_w in mathcalS V(s, w)V(w, g) & textotherwise endcases endaligned]

where $mathcalE$ is the set of edges in the environment’s transition graph, and $V$ is the value function associated with the sparse reward $r(s, g) = 1(s = g)$. Intuitively, this means that we can update the value of $V(s, g)$ using two “smaller” values: $V(s, w)$ and $V(w, g)$, provided that $w$ is the optimal “midpoint” (subgoal) on the shortest path. This is exactly the divide-and-conquer value update rule that we were looking for!

The problem

However, there’s one problem here. The issue is that it’s unclear how to choose the optimal subgoal $w$ in practice. In tabular settings, we can simply enumerate all states to find the optimal $w$ (this is essentially the Floyd-Warshall shortest path algorithm). But in continuous environments with large state spaces, we can’t do this. Basically, this is why previous works have struggled to scale up divide-and-conquer value learning, even though this idea has been around for decades (in fact, it dates back to the very first work in goal-conditioned RL by Kaelbling (1993) – see our paper for a further discussion of related works). The main contribution of our work is a practical solution to this issue.

The solution

Here’s our key idea: we restrict the search space of $w$ to the states that appear in the dataset, specifically, those that lie between $s$ and $g$ in the dataset trajectory. Also, instead of searching for the optimal $textargmax_w$, we compute a “soft” $textargmax$ using expectile regression. Namely, we minimize the following loss:

[beginaligned mathbbEleft[ell^2_kappa (V(s_i, s_j) – barV(s_i, s_k) barV(s_k, s_j))right], endaligned]

where $barV$ is the target value network, $ell^2_kappa$ is the expectile loss with an expectile $kappa$, and the expectation is taken over all $(s_i, s_k, s_j)$ tuples with $i leq k leq j$ in a randomly sampled dataset trajectory.

This has two benefits. First, we don’t need to search over the entire state space. Second, we prevent value overestimation from the $max$ operator by instead using the “softer” expectile regression. We call this algorithm Transitive RL (TRL). Check out our paper for more details and further discussions!

Does it work well?

Your browser does not support the video tag. humanoidmaze

Your browser does not support the video tag. puzzle

To see whether our method scales well to complex tasks, we directly evaluated TRL on some of the most challenging tasks in OGBench, a benchmark for offline goal-conditioned RL. We mainly used the hardest versions of humanoidmaze and puzzle tasks with large, 1B-sized datasets. These tasks are highly challenging: they require performing combinatorially complex skills across up to 3,000 environment steps.

TRL achieves the best performance on highly challenging, long-horizon tasks.

The results are quite exciting! Compared to many strong baselines across different categories (TD, MC, quasimetric learning, etc.), TRL achieves the best performance on most tasks.

TRL matches the best, individually tuned TD-$n$, without needing to set $boldsymboln$.

This is my favorite plot. We compared TRL with $n$-step TD learning with different values of $n$, from $1$ (pure TD) to $infty$ (pure MC). The result is really nice. TRL matches the best TD-$n$ on all tasks, without needing to set $boldsymboln$! This is exactly what we wanted from the divide-and-conquer paradigm. By recursively splitting a trajectory into smaller ones, it can naturally handle long horizons, without having to arbitrarily choose the length of trajectory chunks.

The paper has a lot of additional experiments, analyses, and ablations. If you’re interested, check out our paper!

What’s next?

In this post, I shared some promising results from our new divide-and-conquer value learning algorithm, Transitive RL. This is just the beginning of the journey. There are many open questions and exciting directions to explore:

Perhaps the most important question is how to extend TRL to regular, reward-based RL tasks beyond goal-conditioned RL. Would regular RL have a similar divide-and-conquer structure that we can exploit? I’m quite optimistic about this, given that it is possible to convert any reward-based RL task to a goal-conditioned one at least in theory (see page 40 of this book).

Another important challenge is to deal with stochastic environments. The current version of TRL assumes deterministic dynamics, but many real-world environments are stochastic, mainly due to partial observability. For this, “stochastic” triangle inequalities might provide some hints.

Practically, I think there is still a lot of room to further improve TRL. For example, we can find better ways to choose subgoal candidates (beyond the ones from the same trajectory), further reduce hyperparameters, further stabilize training, and simplify the algorithm even more.

In general, I’m really excited about the potential of the divide-and-conquer paradigm. I still think one of the most important problems in RL (and even in machine learning) is to find a scalable off-policy RL algorithm. I don’t know what the final solution will look like, but I do think divide and conquer, or recursive decision-making in general, is one of the strongest candidates toward this holy grail (by the way, I think the other strong contenders are (1) model-based RL and (2) TD learning with some “magic” tricks). Indeed, several recent works in other fields have shown the promise of recursion and divide-and-conquer strategies, such as shortcut models, log-linear attention, and recursive language models (and of course, classic algorithms like quicksort, segment trees, FFT, and so on). I hope to see more exciting progress in scalable off-policy RL in the near future!

Acknowledgments

I’d like to thank Kevin and Sergey for their helpful feedback on this post.

New Post has been published on https://thedigitalinsider.com/the-friday-roundup-movavi-video-editor-2026-released/

The Friday Roundup - Movavi Video Editor 2026 Released

New Post has been published on https://thedigitalinsider.com/should-freelancers-advertise-their-pricing-speckyboy/

Should Freelancers Advertise Their Pricing? — Speckyboy

Freelancers can choose their own business policies. We can determine how we work, when we work, and how much we charge. That last one can be difficult, to say the least.

Pricing has confounded many a small business owner. Choosing what to charge for your service is only one part of the equation, however. You must also decide how to communicate those figures with others.

Here’s a common challenge: Should you have a pricing page on your portfolio website? This information is a staple for many industries. For instance, everyone expects to know the price of a cup of coffee at their neighborhood shop. People looking for a salon will want to see a price list, as will those booking a cleaning service.

But creative fields like web design are different. We don’t traditionally sell one-size-fits-all commodities, although the industry is shifting. Even so, there are endless ways to build a website. Website maintenance pricing can also vary based on multiple factors.

However, that hasn’t stopped some web professionals from advertising their pricing. That alone makes it a topic worth discussing. An ever-changing industry and evolving client expectations also play a role.

With that, here’s a look at the good and bad of upfront pricing.

Pricing Doesn’t Need to Be Exact

Odds are that no two types of web projects will be priced the same. For example, a five-page brochure website costs less than a 100-site WordPress Multisite network. Adjacent services, such as web hosting, also depend on the project’s size and scope.

That’s a challenge for all but the narrowest of niches. As such, it seems nearly impossible to advertise exact pricing. Perhaps that’s not such a big deal.

There’s something to be said for advertising “ballpark” figures, ones that provide a price range. For one, it helps to weed out low-budget clients. Think of all the time you’ll save!

A price range, or even a “starting at” price, sets the right expectations. Potential clients will have a better idea of costs, while you’ll benefit from a buffer to work within. Just be sure to set a starting price that offers a comfortable profit margin.

In all, transparency is a good thing for your business. This is one way to demonstrate your commitment to being open with clients.

The Pitfalls of Publishing Your Prices

There are a few notable downsides to publicly sharing your pricing. For one, it’s information that competitors can use to their advantage.

If you’re charging $5,000 for a WooCommerce build, then another agency could undercut your price. Even a small discount could cause a client to go with your rival. From there, it becomes a race to the bottom. You might lower your price until there’s little margin to work with.

Public pricing may also complicate things when working on a project estimate. A client may not understand why their estimate is higher than what you advertised. That requires a delicate conversation about the factors involved.

The alternative is to eat the extra cost and hope that things even out. Some projects may require less work, while others take more resources. That’s a dangerous way to live, as you might be leaving money on the table.

Client perception is also a concern. There’s a risk to being seen as too expensive or too cheap. For instance, lower pricing may make some clients see you as inexperienced or low quality. It’s not fair, but people have been known to make snap judgments.

So, if you’re going to publish your prices, be aware of the potential pitfalls. Perhaps they’re worth it, or not. That’s for you to decide.

Is It the Right Move for Your Web Design Business?

The decision to publish your pricing comes down to your individual goals. For instance, making them public might be a way to advertise a new service or a flash sale. That won’t apply to every business, however.

Before you hit the “publish” button, ask yourself the following questions:

How does pricing fit into my overall marketing strategy?

Will it give me an advantage in attracting new clients?

What are the potential downsides?

The important thing is to know why you’re publishing your pricing. Doing so should serve a purpose, even if it’s a temporary experiment. You’ll also want to acknowledge the risks involved.

In all, upfront pricing is still a difficult subject for freelancers. There’s no universally right or wrong answer.

Our best advice is to consider your options carefully and go with what makes sense for your business.

WordPress.com vs. WordPress.org – What’s the difference?

We get this question all the time, and we’re happy to help.

WordPress.org is the most powerful website building software on the web. You will need to find a hosting provider if you want that site online.

WordPress.com is our preferred hosting provider for medium-large traffic websites.

If you want to know why WordPress.com is our preferred host for ambitious passion projects and large website projects, read our review:

Related Topics

Written by Eric Karkovack

Eric Karkovack is a web designer and WordPress expert with over two decades of experience. You can visit his business site here. He recently started a writing service for WordPress products: WP Product Writeup. He also has an opinion on just about every subject. You can follow his rants on Bluesky @karks.com.

New Post has been published on https://thedigitalinsider.com/epiphan-ec20-ptz-camera-with-pearl-system-bundles/

Epiphan EC20 PTZ Camera with Pearl System Bundles

NDI November is a month-long, free, virtual event series dedicated to exploring the power of NDI technology for modern live production. Throughout November 2025, we’ll bring together industry experts, vendors, dealers, and system integrators to deliver an in-depth look at NDI workflows and innovations shaping the future of IP-based production. Join Epiphan on November 6th at 1pm EST to see the new EC20 PTZ Camera and Pearl System Bundles Register for FREE now and don’t miss this ultimate NDI event!

Epiphan has just launched the new EC20 PTZ Camera, along with exclusive bundles that pair it perfectly with Pearl Systems for seamless integration and powerful performance. Contact Videoguys at 800-323-2325 for free tech advice on which bundle is right for you!

4K60p

20x Optical Zoom

NDI|HX2, SRT, TCP/IP,

HTTP, RTSP, RTMP(s) Outputs

1/1.8 inches CMOS Sensor

60° field of view

$1,899.00

Event Streaming Bundle A compact all-in-one kit with a PTZ camera and encoder that makes it easy to stream and record professional quality events, from corporate town halls to live productions, without the complexity of a full control room.

1x EC20

1x Pearl Nano

$3,419.00

Lecture Capture Bundles Higher-ed AV/IT professionals looking for a plugand-play video capture solution that’s reliable, easy to deploy and manage, works with their existing workflows, and can scale across campus.

$4,729.00

Studio Bundles Corporate, media, or podcasting organizations looking for an automated or self-serve video studio that can be operated by anyone to get reliable recorded or live video content

$5,179.00