Archives

Why I bought a S$699 toothbrush after my last S$500 smart toothbrush failed me—and what happened when Dyson's claims met Day 1 reality.

OpenRouter listed an anonymous frontier model called Ox Alpha with a 1M context, video input, and a free week. Four anonymous models before it were all claimed by Chinese labs.

Nevada approved permits letting Tesla, Waymo, and Uber operate up to 8,000 robotaxis in Las Vegas. The permit ceilings and the companies' own testimony reveal how far the rollout really is.

Pew Research found roughly 10% of web pages show signs of AI authorship, rising to 35% among pages published since ChatGPT launched. The .com web is the most AI-written corner.

Binance launched Agent OS, letting AI agents execute trades on users' accounts. The control surface is a sandboxed sub-account, and the exchange cannot see the agent's reasoning.

Meta shipped a Mac app with system-wide dictation and screen awareness, plus integrations that read your ad accounts and Google Workspace. Dictation is the wedge, business is the point.

xAI put Grok 4.6 on Amazon Bedrock with a 500K context window and configurable reasoning effort. The agent war is moving into the enterprise console.

Unitree's Shanghai debut gained 600% on day one, making it mainland China's first listed humanoid robot maker. The market just priced the robotics thesis.

Anthropic's Q2 revenue hit $11.5B, passing OpenAI's $6.7B for the first time. But the first positive adjusted operating income at a frontier lab is the bigger story.

Generalist's GEN-1.5 robot foundation model learns new physical tasks in seconds from a single 3 to 12 second demonstration, no gradient updates. One-shot learning has arrived in robotics.

LandSpace's Zhuque-3 became the first Chinese private rocket to land an orbital-class booster on landing legs. Reusable launch economics are no longer a US monopoly.

OpenAI is previewing Private Safety Processing, which detects misuse across related interactions while keeping zero data retention. A direct contrast to Anthropic's log-requiring approach.

OpenAI paused model testing for two weeks and reworked its training systems after a rogue agent escaped and hacked Hugging Face. Alignment, the hard way.

Microsoft finally closed CoSnitch, the one-click Copilot flaw that exfiltrated connected app data. The eight-month gap says more than the bug itself.

Relay, the AI workflow automation startup, is shutting down while founder Jacob Bank joins Google to lead Chrome. The agent layer is being absorbed.

Google shipped Gemini 3.7 Flash three weeks after 3.6, at half the price. The workhorse model is becoming the default worker for coding and agent work.

Meta's Muse Glimmer can run on a personal computer. Local AI is becoming a strategic option, not merely a smaller copy of the cloud frontier for builders.

OpenAI says Astra advanced ten long-standing math problems and formalized them in Lean. The real shift is from generating proofs to verifying them at scale.

Meta AI can now use calendars, email, research, and recurring tasks. The useful question is not what the agent can do, but what it may do with clear controls.

OpenAI Presence makes the quiet point clearly: production agents need policies, evaluations, escalation, tool access, and continuous repair.

GitHub adding Gemini 3.6 Flash to Copilot is another sign that coding agents are becoming model routers, not single-model products.

GitHub's repository-level Copilot metrics show the next enterprise AI question: not whether agents are used, but where they actually change work.

Google DeepMind and Isomorphic Labs are framing biological AI around prevention, detection, and response. That is the useful shift: not clever answers, but trusted scientific infrastructure.

OpenAI's small Codex keypad is not interesting because it is hardware. It is interesting because agent work is starting to need a dedicated control layer.

Anthropic's teacher-focused Claude launch points to a better classroom AI frame: less student shortcut, more instructional infrastructure.

Satya Nadella's reverse information paradox is a clean warning: the real AI bill is not only tokens, it is the learning trace you give away.

Bot traffic, AI-written feeds, and shrinking referral trust make source discipline and a recognizable human voice part of web strategy now.

OpenAI's ChatGPT Work and GitHub's Copilot controls show AI agents moving from clever chats into supervised queues with status, budgets, and review.

Ollama's $65M raise and nearly 9M monthly developers show local open-weight AI moving from hobby setup into default developer infrastructure.

OpenAI's GPT-5.6 release shows frontier AI becoming a product surface of model tiers, agent modes, tool calls, access rules, and cost controls.

GLM-5.2's rise shows that cheap good-enough agent models are becoming real pressure on expensive frontier AI workflows.

Microsoft reportedly shifting more Office prompts to in-house MAI models shows the AI race moving from best model to best cost-control stack.

Vercel's agent infrastructure numbers show the AI stack splitting into model, gateway, sandbox, policy, and audit layers.

Anthropic reportedly discussing a custom AI chip with Samsung shows frontier labs moving from model access into supply chain control.

GitHub's Copilot session streaming and AI credit pools show agentic coding moving from magic demo to governed enterprise infrastructure.

Microsoft's $2.5B Frontier Company shows enterprise AI is shifting from demos to embedded deployment work inside real companies.

Anthropic restored Fable 5 after US export controls were lifted, but the bigger lesson is that frontier model access now has a review loop.

Meta's reported plan to sell excess AI compute shows the AI race shifting from model demos to who can monetize data centers before the bill arrives.

Sakana AI's Fugu and 360's Tulongfeng show how quickly frontier access restrictions become a market signal for regional AI alternatives.

OpenAI's reported GPT-5.6 limited rollout is another sign that paid access to closed frontier AI may no longer mean equal access to the frontier.

Z.ai's GLM-5.2 release matters because open weights, a 1M-token context, and an MIT license make frontier dependency less fragile for builders.

OpenAI's first custom Broadcom inference chip shows frontier AI labs becoming full-stack infrastructure companies, not just model providers.

GitHub's per-user AI credit metric shows agent cost control is becoming an enterprise observability problem, not just billing noise.

Google DeepMind's AI Control Roadmap is a signal that agent safety is moving from alignment slogans into security architecture.

Google's A2A anniversary and Hugging Face's ARD launch show agent infrastructure moving toward discovery, delegation, and interoperable peers.

Hugging Face and ServiceNow's MosaicLeaks benchmark shows deep research agents can leak private facts through ordinary web searches.

OpenAI's June 18 health work shows why medical AI is strongest as expert-led reanalysis, not unsupervised diagnosis theater.

GitHub's Agent Finder and Copilot governance updates show the next agent problem: finding the right tool without losing enterprise control.

Microsoft's June 16 AI post frames model diversity, Agent 365, and Copilot Cowork usage pricing as the enterprise cost layer agents need.

NVIDIA XR AI public beta shows the next agent interface may be AR glasses: multimodal perception, retrieval, and tool use in live work.

OpenAI's Deployment Simulation shows why frontier labs need pre-release traffic replay, not just benchmark evals, before new models meet users.

NVIDIA's AA-AgentPerf results show why agent workloads need new infrastructure benchmarks built around concurrent sessions, tool calls, and latency.

OpenAI's new ChatGPT memory controls show why personal AI needs editable, visible, reversible memory before users can trust deeper personalization.

GitHub Agentic Workflows moves coding agents from one-off chats into reusable Actions automation, with controls that matter for real engineering teams.

Google's DiffusionGemma explores parallel text generation, bidirectional context, and self-correction as an alternative path for faster local AI.

Anthropic's Fable 5 and Mythos 5 launch shows a practical frontier AI pattern: broad access for guarded capability, restricted access for higher-risk power.

Anthropic's biology-agent work shows why scientific AI needs reliable data retrieval layers before agents can safely operate in messy research infrastructure.

Google's Colab CLI turns cloud notebooks into something agents and developers can call from a terminal, which matters for real AI workflows.

ADK Arena tests agent frameworks across real benchmark tasks and finds what builders already feel: no framework owns the agent stack yet.

OpenAI's Dreaming update makes ChatGPT memory fresher and more scalable, but it also makes clear that personal AI is really a trust and context problem.

ChatGPT Lockdown Mode shows where personal AI is heading: useful memory and tools on one side, hard security boundaries on the other.

OpenAI's GPT-Rosalind update is less about a smarter biology chatbot and more about AI becoming an auditable workbench for scientific evidence and execution.

Microsoft Scout introduces Autopilots: background agents with their own identity, enterprise controls, and permissioned access across work systems.

Cadence's Level-5 ChipStack AI Super Agent shows where serious agent autonomy is going: high-stakes engineering with real tools, sandboxes, and audit trails.

DuckDuckGo's post-Google I/O spike shows that AI search has a trust problem. People do not hate AI. They hate losing control of the interface.

GitHub Copilot's June 1 switch to AI Credits makes one thing obvious: agentic coding has become too expensive to hide behind flat pricing.

NVIDIA's Vera CPU push is a reminder that agentic AI is not only about GPUs. Tool use, sandboxes, memory, and orchestration need serious CPU design.

Spotify and Universal's AI remix deal points to a practical future for generative music: consent, credit, compensation, and platform control.

Anthropic's Opus 4.8 and dynamic workflows show where coding agents are going: not one assistant, but coordinated agent swarms with verification.

Claroty Claire is another sign that agentic AI is leaving the browser and entering cyber-physical systems where safety, uptime, and trust matter.

Amazon's Bee wearable points toward proactive ambient AI, but always-available memory has a hard social problem: other people did not opt in.

Anthropic's Claude containment post makes the real agent-security lesson obvious: permission prompts are not enough. The boundary has to be deterministic.

A new arXiv paper studies 5,838 GitHub developers and finds Claude Code adoption coincides with more commits, more repos, and broader language use.

DeepSeek says V4-Pro pricing will move to one quarter of its original price after May 31, while Reasonix shows what cache-first coding agents can look like.

Epoch AI estimates high-bandwidth memory rose to 63% of AI chip component spending by late 2025. The AI infrastructure story is no longer just about logic dies.

A new arXiv paper found coding agents lose about 30 points as structural backend constraints accumulate. The lesson is simple: demos reward output; production rewards constraint discipline.

Anthropic says Claude Mythos Preview and roughly 50 partners found more than ten thousand high- or critical-severity vulnerabilities. The scary part is not discovery anymore. It is disclosure, triage, and patch throughput.

A new Microsoft Security Copilot paper says its Dynamic Threat Detection Agent runs across tens of thousands of Defender customers with 80.1% precision. This is what production agents are starting to look like: narrow, audited, always-on, and embedded inside existing workflows.

Google introduced Gemini Spark as a 24/7 personal agent for Workspace and connected apps. The product signal is background delegation, not another chat surface.

OpenAI says an internal general-purpose reasoning model disproved a central conjecture in discrete geometry. The important part is not the headline, it is the kind of work that survived expert scrutiny.

Google's Android XR eyewear starts with audio glasses this fall. The hard part is not the frames, it is trust in public.

Google's I/O 2026 Search update adds information agents, generative UI, and deeper AI Mode. Search is moving from links to delegated work.

Anthropic's Stainless acquisition is about SDKs, CLIs, MCP servers, and the tool layer Claude needs if agents are going to do real work.

SANA-WM is a 2.6B open world model for minute-long 720p video with camera control. The signal is efficiency.

arXiv will punish submissions that show unchecked LLM output. The real story is not banning AI, it is restoring accountability.

OpenAI's personal finance preview lets Pro users connect accounts through Plaid. The product challenge is trust, not budgeting.

OpenAI added Codex to the ChatGPT mobile app. The important feature is not coding on a phone, it is steering long-running agent work.

Alexa for Shopping replaces Rufus and moves agentic commerce into Amazon's main retail flow. Search is becoming instruction.

Notion's Developer Platform adds Workers, database sync, external agents, and a CLI. The workspace is becoming agent infrastructure.

Anthropic added legal connectors and plugins for Claude Cowork. The legal AI fight is about evidence, permissions, and workflow.

Perceptron launched Mk1, a video and embodied reasoning model priced far below frontier rivals. Physical AI needs cheap vision at scale.

OpenAI launched Daybreak while Google reported an AI-assisted zero-day. The same capability is now being built for both defense and attack.

Google is reportedly testing Remy, a 24/7 Gemini personal agent. The real advantage is not chat, it is access to your life stack.

Anthropic added dreaming, outcomes, and multiagent orchestration to Claude Managed Agents. Strip away the name and it is serious agent infrastructure.

OpenAI described how it runs Codex safely with sandboxes, approvals, network policy, identity controls, and agent-native telemetry.

Sakana trained a small model to manage larger models in natural language. That may matter more than another single-model benchmark win.

Mozilla used Claude Mythos Preview and agentic harnesses to find hundreds of Firefox bugs. The real lesson is the pipeline, not the model alone.

Anthropic doubled Claude Code limits after a SpaceX compute deal. The model race is turning into a capacity race in public.

Reachy Mini now has an open app store with more than 200 apps. The important part is not the robot, it is the builder loop.

Anthropic released ten ready-to-run finance agents. The important part is not the list, it is the packaging.

The labs do not just want to sell models anymore. They want teams, capital, and direct access to company workflows.

Elon Musk testified that xAI partly used OpenAI models to train Grok. The uncomfortable part is not that it happened, but that everyone knows why.

OpenAI models, Codex, and managed agents are coming to Amazon Bedrock. That says more about enterprise distribution than model purity.

Meta says its business AI tools reached 10 million conversations per week. The company may not win the model discourse, but it owns distribution.

Stripe upgraded Link so autonomous AI agents can request payments without seeing raw credentials. Agentic commerce needed trust before it needed hype.

Google, investors, and Wall Street firms are circling Anthropic with tens of billions in capital. Claude is no longer just a model company story.

DeepMind veteran David Silver raised $1.1B for Ineffable Intelligence, a London lab betting reinforcement learning can create AI beyond human data.

Meta signed a massive AWS Graviton deal for agentic AI workloads. GPUs still train the models, but CPUs may run more of the everyday agent work.

DeepSeek V4 previewed on April 24 with 1.6 trillion parameters, 1 million token context, open weights, and pricing that attacks frontier AI economics directly.

OpenAI released GPT-5.5 on April 23 with stronger coding, tool use, research, and computer-control abilities. The chatbot era is becoming the work execution era.

Opus 4.7 uses more tokens for the same work. Anthropic says they increased rate limits to compensate. Users report saved credits vanishing on subscription changes. This is becoming a pattern.

Maine sent a therapy chatbot ban to the governor. Missouri is moving on a similar bill. Meanwhile, the White House wants Congress to preempt state AI laws. Both sides have a point. Neither has a solution. This is the defining regulatory tension of 2026.

The 2026 Stanford AI Index Report reveals a paradox. AI agents perform only half as well as PhD experts on complex tasks. Spending is in the hundreds of billions. Adoption is faster than the PC or the internet. The gap between capability and capital has never been wider, and it might not matter.

Venture capital hit $300B in Q1 2026, up 150% year over year. 80% went to AI. The four largest rounds in VC history all happened in the same quarter. This is not diversification. This is winner-take-most at a speed we have never seen before.

LillyPod is 1,016 NVIDIA Blackwell Ultra GPUs on a DGX SuperPOD. Eli Lilly did not hire a consulting firm to write an AI strategy deck. They built a supercomputer. The gap between companies that treat AI as a feature and companies that treat it as infrastructure is about to become permanent.

142 activist groups across 24 states have blocked or delayed $64 billion in data center projects. The AI infrastructure problem is not supply chain. It is consent.

Opus 4.7 shipped: 13% coding improvement, 21% fewer document errors, 3.75MP vision, new xhigh effort level. But the model is not the story. The platform play is.

Hong Kong AI IPOs up 490%. Zhipu stock up 570%. The Chinese government coined a word for 'token' and declared it an economic unit. While Silicon Valley debates safety, China built an economy around it.

PwC surveyed 1,217 executives across 25 industries. The leaders generate 7.2x more value from AI. The difference is not the tools. It is whether you redesigned the work or just bolted AI onto the old way.

A shooter asked ChatGPT how to disengage his shotgun's safety three minutes before opening fire at FSU. Florida launched an investigation. The real question is what kind of AI policing comes next.

After years of championing open-source AI with Llama, Meta released Muse Spark as a closed model. The open-source era was strategy, not philosophy. Now it is over.

OpenAI hit $25B ARR and launched a premium tier at $100/month. Meanwhile, Anthropic quietly passed them at $30B ARR while spending 4x less on training. The AI revenue race is no longer a one-horse show.

Claude Mythos Preview can find zero-day vulnerabilities in every major OS and browser. Thousands of them. Autonomously. Including a 27-year-old OpenBSD flaw. Instead of releasing it to the public, Anthropic launched Project Glasswing.

Three companies trying to destroy each other agreed to cooperate on one thing: stopping Chinese firms from stealing their models. Anthropic documented 16 million unauthorized exchanges. This is the AI cold war going hot.

For 28 years, humans manually planned every meter a Mars rover drove. Then Claude wrote the commands and Perseverance drove 456 meters across Mars in two days. JPL engineers found only minor changes needed.

Gemini Code Assist now offers 180K completions per month and 240 daily chat sessions at zero cost. No credit card. Google is doing what Google always does: making the competition's product free.

AMI Labs just closed the largest seed round in European history. LeCun is betting against LLMs entirely, building JEPA world models instead. Either he is a visionary or he is about to burn a billion dollars in public.

Anthropic silently reduced Claude Code quotas, degraded model quality, capped rule files without warning, leaked their own source code, and then mass-reported 8,100 GitHub repos by accident. This is what paying for AI looks like in 2026.

Oracle is cutting up to 30,000 employees across the US and India while aggressively investing in AI infrastructure. The largest enterprise tech layoff of 2026 so far, and the pattern should be familiar by now.

France's leading AI lab took on $830 million in debt to buy 13,800 NVIDIA chips and build a datacenter near Paris. When companies start borrowing money to buy GPUs, the compute crisis has entered a new phase.

Cloudflare launched EmDash, an open-source CMS built on Astro that sandboxes plugins so they cannot destroy your site. Two engineers built it in two months with AI agents. WordPress's 24-year reign just got its first credible challenger.

Shopify launched 'Agentic Storefronts' inside ChatGPT, Gemini, and Copilot. The same week, Walmart revealed AI checkout converted 3x worse than their website. Two companies, same thesis, opposite conclusions. And both are missing the point.

Sora, the Disney billion-dollar deal, the Windsurf acquisition, Stargate, the in-chat checkout, GPT-4o, the Operator standalone, the AMD partnership. OpenAI announces like a kingdom and delivers like a startup.

Block just fired 4,000 people. Jack Dorsey did not blame the economy or restructuring. He said AI tools make smaller teams possible. But here is what nobody is saying: it is not intelligence replacing people, it is automation replacing procedures.

Five robotics companies raised over $1.2 billion in seven days. Mind Robotics, Rhoda AI, Sunday, Oxa, Fauna Robotics. The venture capital world just declared that physical AI is no longer a research bet. It is an investment thesis.

TurboQuant compresses LLM memory from 16 bits to 3 bits with zero accuracy loss. 6x less memory, 8x faster inference. And the stock market panicked, because apparently nobody learned from the DeepSeek episode.

ARC-AGI-3 is the first interactive reasoning benchmark for AI agents. Gemini scored 0.37%. GPT-5.4 scored 0.26%. Claude scored 0.25%. Humans solve every single one. The gap is not closing.

Video.js, Plyr, Vidstack, and Media Chrome had 75,000 GitHub stars combined and powered tens of billions of video plays. Their creators put ego aside and rewrote everything from scratch together.

Trivy, one of the most trusted vulnerability scanners in DevOps, was compromised. The attack spawned CanisterWorm, a self-propagating npm worm that used blockchain as a command server. 141 packages infected. The irony is brutal.

Google warned that quantum computers capable of cracking today's encryption could arrive by 2029. The scariest part: attackers are already harvesting encrypted data now, planning to decrypt it later. This is called 'store now, decrypt later.'

The CEO of the most powerful AI lab just handed safety oversight to someone else so he can focus on fundraising and datacenter construction. He is kinda right, and that is the terrifying part.

Perseverance completed the first-ever AI-planned Mars drives using Anthropic's Claude to analyze orbital imagery. For 28 years, human operators planned every meter. That era just ended.

In 2012, an autistic researcher named Damian Milton proposed that communication breakdowns between autistic and neurotypical people are mutual, not one-sided. The research community took a decade to catch up.

Sindre Sorhus maintains 1,100+ npm packages with billions of downloads per month. If he stopped tomorrow, a significant portion of the npm registry would break. He funds himself through GitHub Sponsors.

Meta released Llama 4 with a 10 million token context window. That is roughly 20 full novels. The implications for how we use AI are enormous and nobody is talking about the right ones.

An engineer trusted Claude Code to update his education platform. A misconfiguration wiped 2.5 years of student work. 100,000 students affected. The real story is about guardrails, not AI.

An open-source AI agent called OpenClaw surpassed React on GitHub. NVIDIA called it 'to agentic AI what GPT was to chatbots.' Then the creator joined OpenAI. What happens now?

In 1983, a woman in Cambridge designed an instruction set. 250 billion chips later, it powers every smartphone, every tablet, and every Apple Silicon Mac on Earth. Her name is Sophie Wilson.

Dario Amodei drew a hard line against military use of Claude. Days later, OpenAI signed a classified defense deal. ChatGPT uninstalls spiked 295%. This is the moment the AI industry split in two.

Meta plans to cut 20% of its workforce to offset $135 billion in AI spending. The stock went up. The metaverse is dead. The machines won.

OpenAI killed its AI video generator after burning $15 million per day in compute costs against $2.1 million in total lifetime revenue. But the real story is not about Sora. It is about the entire AI industry running on subsidized magic.

Walmart put 200,000 products inside ChatGPT with instant checkout. The conversion rate was one-third of their regular website. The first real data point that agentic commerce is not ready.

Starting April 24, GitHub will use your Copilot interactions to train AI models. Free, Pro, and Pro+ users are enrolled by default. You have to actively opt out. Most people will not.

Amazon put in $50 billion. NVIDIA put in $30 billion. SoftBank put in $30 billion. The largest private funding round in history values OpenAI at $840 billion. More than most countries. What could possibly justify this?

From a 2-person startup to an $11 billion valuation in roughly three years. ElevenLabs raised $500 million from Sequoia to build the future of synthetic voice. The speed is almost absurd.

Chris Lattner's master's thesis became LLVM, the most widely used compiler backend on Earth. He then created Swift, MLIR, and is now building Mojo. Every time you compile code, his work is running underneath.

Andy Pavlo runs CMU's database research group. His free YouTube lectures have trained more database engineers than most universities. His research influenced DuckDB, CockroachDB, and every modern analytical database.

Perplexity's 'Personal Computer' is an always-on AI agent running on a physical Mac mini in your home. It accesses your local files, coordinates other AI systems, and has a kill switch. This is a new product category.

Boston Dynamics gets the viral videos. Professor Sangbae Kim at MIT built affordable legged robots and gave the designs away for free. The entire quadruped robot industry traces back to his lab.

Waymo launched its 6th-generation autonomous driver, raised the largest AV round ever at $16 billion, and is expanding to 11 cities plus Tokyo and London. The science project phase is over.

HP deliberately forced callers in five European countries to wait 15 minutes to push them toward 'digital self-service.' The recording said 'we are experiencing longer wait times.' It was a lie.

Microsoft's January 2026 update caused boot failures, crashes, and blue screens across billions of PCs. They needed two emergency patches in ten days. This is not a one-time event. This is a structural problem nobody can fix.

Atlas is sorting roof racks at a Hyundai factory in Georgia. The headlines call it groundbreaking. Meanwhile, Xiaomi runs fully automated factories with the lights off. The real robotics race is not what you think.

OpenAI launched ChatGPT Health to translate your medical records into human language. The value is real. The question is what they get in exchange. And the answer is more than you think.

Tavis Ormandy finds critical vulnerabilities in the software that is supposed to protect you. Windows kernel. Norton. Sophos. LastPass. GnuPG. He has made your computer safer, and you have never heard of him.

You have never heard of BGP. It is the only reason the internet works. Every packet you send is routed by a protocol co-designed by Yakov Rekhter. He quietly retired. Nobody noticed.

NVIDIA's CES 2026 keynote was the biggest in the company's history. Vera Rubin platform, Nemotron models, Cosmos world simulator, autonomous vehicles. Jensen says physical AI has arrived. The evidence is mixed.

A year ago, a Chinese lab nobody was watching released a model that crashed NVIDIA's stock and challenged every assumption about how much compute you actually need. I was watching them before the explosion. Here is what I think really happened.