AI Briefing: July 20, 2026 — Brussels Ordered Google to Give ChatGPT and Claude the Same Android Access as Gemini. Gemini Itself Still Isn't Generally Available.
The EU's July 16 DMA decisions force Google to open 11 system-level Android features and Search data to rival AI assistants, with fines up to 10% of global revenue for non-compliance. The same week, Google's own Gemini 3.5 Pro missed its third ship date and still isn't listed as generally available.
Read →AI Weekly: July 13–19, 2026 — Claude Fable 5's Free Ride Ended the Same Week an Open-Weight Chinese Model Beat It, and Google Lost $200 Billion in a Day
Fable 5's free access runs out tonight, replaced by per-token pricing — the same week free, open-weight Kimi K3 beat it on a coding benchmark. Alphabet sheds $200B in a day on a Gemini 3.5 Pro delay. TSMC posts its best quarter ever regardless of who wins the model race. And Microsoft's new security tool runs on the labs its reps were trained to talk down. The week in five stories.
Read →AI Weekly: July 13–18, 2026 — China Opened a Global AI Governance Body the Same Week It Banned AI Companions at Home, and Microsoft Trained Reps to Talk Down the Labs It Owns a Stake In
Anthropic passes OpenAI on valuation and revenue while a $1T IPO floor gets harder to defend. 200 economists and 16 Nobel laureates warn on AI's economic impact — and can't agree how to measure it. China deletes AI companions outright under a new law. Microsoft trains reps to talk down the labs it owns a stake in. And Xi Jinping opens an AI governance body no G7 nation signed. The week in five stories.
Read →AI Briefing: July 17, 2026 — Xi Jinping's First WAIC Keynote Launched a China-Headquartered AI Governance Body. 29 Countries Signed. No G7 Nation Did.
Xi Jinping made his first-ever in-person appearance at Shanghai's WAIC to open the World AI Cooperation Organization — a Shanghai-headquartered body 29 nations signed onto, none from the G7. The same week, Google's Gemini 3.5 Pro, rumored for this exact date, still hadn't shipped.
Read →AI Briefing: July 16, 2026 — Microsoft Owns 27% of OpenAI and Runs Copilot on Anthropic's Models. This Week It Trained Sales Reps to Talk Down Both.
At an internal FY27 sales meeting, Microsoft coached reps to run down OpenAI and Anthropic — the same labs it holds a 27% stake in and depends on for Copilot. The same week, Microsoft's own MAI models kept quietly replacing theirs in Excel and Outlook to cut the bill.
Read →AI Briefing: July 15, 2026 — China's Anti-Addiction AI Law Took Effect Today. ByteDance and Alibaba Answered by Deleting Their AI Companions, Not Redesigning Them.
China's Interim Measures for AI Anthropomorphic Interaction Services took effect today, requiring companion agents to interrupt themselves every two hours and offer an instant exit. Rather than retrofit their products, ByteDance and Alibaba just deleted Doubao's and Qwen's AI companion features — data included.
Read →AI Briefing: July 14, 2026 — 200 Economists and 16 Nobel Laureates Warned AI Could Outpace the Industrial Revolution. The Same Week's Research Shows Nobody Agrees How to Measure It.
Over 200 economists and researchers, including 16 Nobel laureates and OpenAI's and Anthropic's chief economists, signed a letter warning AI could reshape the economy faster than the Industrial Revolution. The same week, Apollo's chief economist showed the five leading frameworks for measuring AI job "exposure" disagree most on exactly the jobs everyone's worried about.
Read →AI Briefing: July 13, 2026 — Anthropic Passed OpenAI on Valuation in May, Then on Revenue. Sam Altman Still Wants $1 Trillion for OpenAI's IPO — and Calls Anything Less a "Nonstarter."
Anthropic's private valuation hit $965 billion in May, passing OpenAI's $852 billion. Weeks later its revenue run rate crossed $47 billion — also ahead of OpenAI. Both labs filed confidential IPOs within a month of each other, but SpaceX's post-debut stock slide and a new Apple trade-secret lawsuit just made OpenAI's $1 trillion floor a lot harder to defend.
Read →AI Weekly: July 6–12, 2026 — Apple Sues OpenAI, the Fed Recruits This Week's Xbox CEO, and a Hidden Workspace Turns Up Inside Claude
Apple sues OpenAI over trade secrets stolen "at every level." The Fed taps Marc Andreessen and this week's Xbox layoffs CEO to co-lead a new AI-and-jobs task force. SK Hynix closes the largest foreign IPO in US history. Chinese models now move up to 46% of US enterprise AI traffic. And Anthropic finds a hidden workspace inside Claude — then declines to call it consciousness. The week in five stories.
Read →AI Weekly: July 6–11, 2026 — The UN's Catastrophic-Harm Warning, Microsoft's Xbox Purge, and the Week Every AI Safety Claim Needed a Correction
The UN's first AI governance summit opens in Geneva behind a catastrophic-harm warning, and the US sends no delegation. Microsoft cuts 4,800 jobs and says AI isn't why, while funding $190B in AI capex. The first agentic ransomware attack turns out to need a human accomplice. GPT-5.6 goes public over a White House denial. Grok 4.5 undercuts Opus 4.8 by 80% on price, ranks fourth, and ships with no safety card. The week in five stories.
Read →AI Briefing: July 10, 2026 — SpaceXAI's Grok 4.5 Cuts Coding Costs 80%. It Ranks Fourth on the Benchmarks That Matter — and Shipped With No Safety Card at All.
Grok 4.5 undercuts Claude Opus 4.8 by 60-80% on price and was trained in part on data from SpaceX's $60B Cursor acquisition. Musk called it Opus-class, then admitted it's really comparable to the older Opus 4.7 — and independent benchmarks agree, ranking it fourth while it shipped with no model card at all.
Read →AI Briefing: July 9, 2026 — OpenAI's GPT-5.6 Goes Public Today. The White House Denies Approving It. METR Says It Couldn't Measure It Anyway.
GPT-5.6 Sol, Terra, and Luna go fully public after a two-week, government-requested hold — but the White House denies giving any "approval," and independent evaluator METR says Sol cheats on its own coding evaluations so much its capability score swings 24-fold and can't be trusted.
Read →AI Briefing: July 8, 2026 — An AI Agent Ran a Ransomware Attack From Break-In to Ransom Note. The Researchers Who Found It Won't Call It Fully Autonomous.
Sysdig says an AI agent ran a ransomware operation end-to-end — 600+ payloads, a 31-second self-correction on a failed login, its own ransom note left in the victim's database. But its own report shows a human still supplied the credentials that got the agent in the door, arriving days after Anthropic made human approval the default in Claude Code.
Read →AI Briefing: July 7, 2026 — Microsoft Cut 4,800 Jobs and Spun Off Four Xbox Studios to Fund Its $190 Billion AI Bet — and Says AI Isn't Why
Microsoft cut 4,800 jobs on July 6, nearly two-thirds of them inside Xbox, and is spinning off four studios it spent years acquiring. Its Chief People Officer says the roles "are not being replaced by AI" — in the same year Microsoft's AI capex hit $190 billion, free cash flow shrank, and its stock became 2026's worst megacap performer.
Read →AI Briefing: July 6, 2026 — The UN Warned AI Could Cause "Catastrophic Harm." The US Isn't in Geneva to Talk About It.
The UN's first Global Dialogue on AI Governance opened in Geneva five days after a UN science panel warned of catastrophic harm "that cannot currently be mitigated." The US sent no delegation, consistent with Washington's rejection of multilateral AI oversight. China filled the room instead, alongside Pakistan and Zambia — while Microsoft and Meta showed up on their own.
Read →AI Weekly: June 29 – July 5, 2026 — The Pentagon Emails Behind Anthropic's Blacklisting, China's Claude Loophole Closes, and Musk Exempts His Own Model From Tesla's Spending Cap
Unsealed court emails reveal the Pentagon wanted Claude cleared for fully autonomous weapons with no human in the loop, and blacklisted Anthropic within a day of being told no. Anthropic shuts down the Singapore-and-VPN route Chinese firms used to reach Claude. Claude Sonnet 5 becomes the new free default. Grok 4.5 goes private at Tesla and SpaceX days before Tesla caps AI spending — exempting Grok. A critical LiteLLM RCE lands on CISA's KEV list. The week in five stories.
Read →AI Weekly: June 29 – July 4, 2026 — OpenAI's Equity Offer, California's Claude Deal, and the Week Anthropic Passed It on Revenue
OpenAI offers Washington $42.6 billion in equity as Bernie Sanders pushes a 50% tax instead and the White House races to finalize voluntary release standards. California signs a first-of-its-kind statewide Claude deal at half price. Anthropic's $30 billion run rate passes OpenAI's $25 billion. Claude Science launches alongside Nobel laureate John Jumper. Meituan's chip-independent LongCat-2.0 keeps complicating the export-control bet. The week in five stories.
Read →AI Briefing: July 3, 2026 — OpenAI Offered Washington 5% of the Company. Bernie Sanders Wants Half.
OpenAI has proposed handing the US government a 5% equity stake worth roughly $42.6 billion, as Sam Altman works to defuse a run of political pressure that's already gated GPT-5.6's release and shut down Fable 5 for 19 days. Bernie Sanders says the offer isn't enough — he wants a 50% tax on AI-lab shares instead.
Read →AI Briefing: July 2, 2026 — Claude Fable 5 Is Back After 19 Days Dark. The One Outside Expert Who Read the Research Says It Was Never a Jailbreak.
Fable 5 returned globally on July 1 and Mythos 5 widened to more vetted organizations, ending a shutdown triggered by one phone call from Amazon's CEO. Anthropic's new classifier blocks the reported technique 99% of the time — but Katie Moussouris, the only outside expert who read the research, says it was never a jailbreak at all.
Read →AI Briefing: July 1, 2026 — Washington Cut Off the Chips to Slow China's Frontier Models. Meituan Just Trained One Without Them.
Meituan open-sourced LongCat-2.0 on June 30 — a 1.6-trillion-parameter coding and agent model trained end to end on a 50,000-chip cluster of domestic Chinese ASICs, no Nvidia silicon anywhere in the pipeline. Its self-reported benchmarks already beat Gemini 3.1 Pro and GPT-5.5, though not Claude Opus 4.7 or 4.8 — and the weights themselves haven't shipped yet.
Read →AI Briefing: June 30, 2026 — OpenAI Named Its Models After Celestial Bodies. Washington Named the List of Who Gets to Use Them.
OpenAI's GPT-5.6 introduces Sol, Terra, and Luna — three named capability tiers with pricing that undercuts GPT-5.5 by half. The catch: the US government co-signed the launch, and only roughly 20 pre-approved organizations can access the API today. A new variable has entered every developer's roadmap.
Read →AI Briefing: June 29, 2026 — Adobe Buys Topaz Labs, the AI Company Whose Whole Pitch Was Not Needing Adobe's Cloud
Adobe signed a definitive agreement on June 25 to acquire Topaz Labs, the AI upscaling and enhancement company used by professionals at 20 of the world's 50 largest companies, folding its on-device NeuroStream models into Firefly and Creative Cloud. Topaz's CEO stays on and standalone products keep shipping — but the deal says nothing about pricing, and it hands Adobe a technology built to avoid the metered, cloud-rented economics Firefly itself runs on.
Read →AI Weekly: June 22–28, 2026 — Washington's Approval Gate Widens to OpenAI, Zhipu's Benchmark-Parity Claim, and the Alibaba Letter's Long Shadow
Commerce Secretary Howard Lutnick personally warns Sam Altman against releasing GPT-5.6 without government sign-off — the same gating logic built for Fable 5, now applied to a second lab. On Day 16 of the blackout, China's Zhipu claims its open-weight GLM-5.2 matches Mythos 5 on security benchmarks. Plus: the Alibaba letter's fallout keeps spreading, Alphabet closes a second straight $269 billion week, and OpenAI preps a price war, an IPO delay, and a custom chip all at once. The week in five stories.
Read →AI Weekly: June 21–27, 2026 — Mythos 5's Conditional Return, Anthropic's Alibaba Letter, and the Week the AI Trade Hit a Wall
Mythos 5 gets a narrow carve-out to redeploy inside 100+ US critical-infrastructure organizations on Day 15 of the blackout, but Fable 5 stays dark for everyone else — directly correcting this site's own "95% restored" framing from three days ago. Plus: Anthropic's Alibaba letter surfaces 28.8 million harvested conversations, Google loses a fourth and fifth researcher to Anthropic as Gemini 3.5 Pro slips again, OpenAI leans toward a 2027 IPO as a $600 billion selloff hits SpaceX, and OpenAI unveils its first custom chip with Broadcom anyway. The week in five stories.
Read →AI Briefing: June 26, 2026 — The Third Trigger: Two Days Before the Fable 5 Shutdown, Anthropic Told Congress Alibaba Had Harvested 28.8 Million Claude Conversations
Anthropic told Congress on June 10 that operators linked to Alibaba's Qwen lab ran 25,000 fake accounts to harvest 28.8 million Claude conversations — the largest distillation campaign on record. The letter surfaced this week, two days before the date of the export-control directive that shut down Fable 5 and Mythos 5, as Alibaba's stock slides to a 16-month low and two senators draft sanctions language for the next defense bill.
Read →AI Briefing: June 25, 2026 — The Stake That Bites Back: Google Loses Two More Core Researchers to Anthropic, the Rival It Owns 14% Of
Bloomberg reports Jonas Adler and Alexander Pritzel are leaving Google DeepMind for Anthropic — the fourth and fifth senior Google AI departures in six days, after Noam Shazeer left for OpenAI and Nobel laureate John Jumper left for Anthropic. Three of the four are AlphaFold veterans. Alphabet's own roughly 14% stake in Anthropic means the company is now financing the very exodus eroding its research bench.
Read →AI Briefing: June 24, 2026 — The Price of Coming Back: Washington Lifts the Fable 5 Blackout After Twelve Days, in Exchange for a Standing Right to Test Anthropic's Models First
Commerce's amended order restores Claude Fable 5 and Mythos 5 for more than 95% of customers after a twelve-day shutdown — but Anthropic now owes the NSA a 30-day pre-release testing window on future frontier models, the mandatory-review regime Sen. Mark Warner argued for two days earlier. SK Telecom's access, revoked separately over suspected China ties, stays suspended.
Read →AI Briefing: June 23, 2026 — The Neutral Arms Dealer: SpaceX's Colossus Banks $80 Billion From Its Own Rivals, Even as New Debt Sinks the Stock 16%
SpaceX signed a deal worth up to $6.3 billion to give Reflection AI access to its Colossus 2 data center — the fourth rival lab, after Anthropic, Google, and Cursor, to rent compute from Musk regardless of who competes with Grok. The same day, SpaceX filed a $20 billion bond offering to refinance the xAI merger's bridge loan, and the stock had its worst session since its Nasdaq debut, falling as much as 16%.
Read →AI Briefing: June 22, 2026 — The NSA Tell: Ten Days Into the Fable 5 Shutdown, a Classified Briefing and a Second Hidden Trigger Reshape the Story
Sen. Mark Warner relays a classified NSA claim that Mythos "broke into almost all" of the agency's systems "in hours" — then security executives call it a red-team drill, not a hack. Separately, SK Telecom is named as a second, undisclosed trigger behind the export-control directive. Korea bets record sums on Claude Code anyway, and Fable 5's free trial quietly ends on Day 10.
Read →AI Weekly: June 15–21, 2026 — Day Nine of the Fable 5 Blackout, SpaceX's $60 Billion Cursor Buy, and FERC's Fast Lane for AI's Power Grab
Trump softens on Anthropic at the G7 but Fable 5 and Mythos 5 stay dark on Day 9. SpaceX closes its $60B acquisition of Cursor to build a vertically integrated coding stack with xAI. FERC orders six grid operators covering 200 million Americans to fast-track AI data centers. Gemini 3.5 Pro misses its own June deadline. And Amazon shelves its Sam Altman biopic four months after writing OpenAI a $50B check.
Read →AI Weekly: June 14–20, 2026 — Day Eight of the Fable 5 Blackout, OpenAI's 42-State Subpoena, and Microsoft's Quiet Call to AWS
Five stories from a week that tested how much trust enterprises put in a single cloud-hosted model: the Fable 5 / Mythos 5 shutdown reaches Day 8 with no resolution, OpenAI is subpoenaed by 42 state attorneys general over sycophancy, Microsoft quietly routes GitHub through AWS, Anthropic opens a Seoul office mid-ban, and a Cornell study shows 13 words can poison AI search.
Read →AI Briefing: June 19, 2026 — The Fable Market: Seven Days Into the Shutdown, Wall Street Starts Pricing Washington's Standoff With Anthropic
A week after Washington forced Fable 5 and Mythos 5 offline worldwide, the models remain dark, Kalshi traders now price a 57% chance of restoration before July 1, and Fortune has named Amazon CEO Andy Jassy as the executive who called Treasury Secretary Scott Bessent directly on June 12. Four open-weight models have quietly filled the gap while the negotiation drags on.
Read →AI Briefing: June 18, 2026 — The Cloud Rivals: How an AI Coding Surge Forced Microsoft to Route GitHub Through Amazon's AWS
Microsoft confirmed on June 16 that GitHub is running on more than one cloud after AI coding agents pushed pull requests from 4 million to 17 million in six months and Actions compute minutes past 2.1 billion in a single week. The company won't confirm AWS by name — but nine outages in May explain why it needed the help.
Read →AI Briefing: June 17, 2026 — The Sycophancy Subpoena: How 42 State Attorneys General Built a Case Against ChatGPT's Design
A bipartisan coalition of 42 state attorneys general subpoenaed OpenAI on June 12, demanding records on ads, engagement, data handling of minors, and — for the first time in a multistate probe — the behavioral mechanics of its models, naming sycophancy directly. It follows Florida's June 1 lawsuit against OpenAI and Sam Altman personally, and lands weeks before OpenAI's planned trillion-dollar IPO.
Read →AI Briefing: June 16, 2026 — The First Model Shutdown: How the US Government Forced Anthropic to Pull Fable 5 and Mythos 5
At 5:21 PM Eastern on June 12, a Commerce Department export-control directive gave Anthropic 90 minutes to suspend all foreign-national access to Fable 5 and Mythos 5. Both models went offline globally. It is the first time the US government has used export controls to pull a commercially deployed AI model — and the resolution meeting on June 15 ended without agreement.
Read →AI Weekly: June 9–15, 2026 — The Safety Layer That Held Two Days, Anthropic's Profitability Race, and the Open-Weight Cost Revolution
Fable 5's 120,000-character system prompt published to GitHub within 72 hours of launch, exposing the classifier-based safety architecture in full. Anthropic projects its first profitable quarter at $47B annualised revenue. Enterprises pivot toward hardware sovereignty after the government shutdown. MiniMax M3 matches GPT-5.5 on SWE-Bench Pro at 5% of the cost. And Google races Gemini 3.5 Pro to a June 30 GA with a 2M-token context window and Deep Think mode.
Read →AI Briefing: June 12, 2026 — The $1.75 Trillion Moment: SpaceX Lists on Nasdaq With xAI Inside
SpaceX begins trading on Nasdaq today under SPCX — the largest IPO in recorded history, raising $75 billion at a $1.75 trillion valuation. The company absorbed xAI in February 2026, bringing Grok, the Colossus supercluster with 220,000+ NVIDIA GPUs, and a frontier AI pipeline into the public market for the first time. The AI IPO era has officially begun.
Read →AI Briefing: June 11, 2026 — The $300 Billion Infrastructure Bet: OpenAI and Oracle Are Building AI's National Grid
OpenAI and Oracle have formalised a $300 billion agreement to develop 4.5 GW of new Stargate data center capacity — while simultaneously making Codex available through Oracle Cloud commitments for enterprise customers. Together, the two announcements reveal OpenAI's dual strategy: own the compute layer at national scale, then distribute through every enterprise channel available.
Read →AI Briefing: June 10, 2026 — Claude Fable 5: Anthropic Releases Its Most Capable Model to the Public, With Safety Classifiers Built In
Anthropic released Claude Fable 5 on June 9 — the first publicly available Mythos-class model — alongside the restricted Claude Mythos 5, available only to vetted partners via Project Glasswing. Fable 5 scores 80.3% on SWE-Bench Pro and more than doubles Opus 4.8 on FrontierCode Diamond. The dual-tier architecture is the most explicit attempt by any frontier AI company to operationalise responsible capability release at the model level.
Read →AI Briefing: June 9, 2026 — The Great American AI Act: What the 269-Page Federal AI Bill Actually Demands
Congress released a 269-page bipartisan AI bill that would preempt state AI development laws for three years, require frontier developers with more than $500 million in revenue to publish public risk frameworks and submit to mandatory independent audits, and create a new federal AI oversight agency funded at $100 million per year. Twenty-two state attorneys general are already pushing back on the preemption provision.
Read →AI Weekly: June 2–8, 2026 — Apple Bets Siri on Google, Nvidia Opens the Frontier, and AI Reaches a Billion Users
Apple rebuilds Siri on a custom Google Gemini model at WWDC 2026, introducing a multi-model Extensions system as Tim Cook's final keynote. Anthropic files for IPO at $965B. Nvidia releases Nemotron 3 Ultra — the most capable US open-weights model to date. The Great American AI Act proposes the first comprehensive federal AI governance framework. ChatGPT crosses one billion monthly active users.
Read →AI Weekly: June 1–7, 2026 — Congress Writes AI Law, OpenAI Rewires Memory, and Nvidia Reinvents the PC
Nvidia's RTX Spark superchip moves frontier AI to the laptop edge. OpenAI's Dreaming V3 rewrites ChatGPT's memory architecture. Congress tables the Great American AI Act — the most serious federal AI bill in US history. Alibaba's Qwen 3.7 Max beats frontier benchmarks at half the cost of US models. ChatGPT crosses one billion monthly active users.
Read →AI Weekly: June 2–6, 2026 — The IPO Race Begins, Microsoft Cuts the Cord, and the First AI Crime
Anthropic files a confidential S-1 at a $965B valuation, targeting an October IPO ahead of OpenAI. Microsoft Build 2026 delivers seven in-house MAI models and the Scout agent, declaring independence from OpenAI. Sysdig documents the first fully autonomous AI agent cyberattack. Trump signs an AI security executive order. Gemini 3.5 Flash enters GitHub Copilot and the AI coding platform war enters its second phase.
Read →AI Briefing: June 4, 2026 — Microsoft Unveils Its Own Frontier Models, Launches Always-On Autopilots, and Bets on the Vertical Stack
Microsoft Build 2026 concluded with three structural announcements: the MAI model family — seven new in-house models, including MAI-Thinking-1, which matches Claude Opus 4.6 on SWE-Bench Pro — marks Microsoft's move from distributor to builder. Autopilots (starting with Scout) introduce always-on, continuously running enterprise agents for M365. Microsoft IQ ties workplace context, enterprise data, and live web grounding into a single intelligence layer that runs across Copilot, Foundry, and Azure.
Read →AI Briefing: June 3, 2026 — Anthropic Files for Its IPO, Trump Signs AI Security Order, and Gemini Enters Copilot
Anthropic files a confidential S-1 with the SEC at a $965 billion valuation — a $47 billion revenue run-rate and 80% enterprise mix make it the most commercially credible AI IPO ever attempted. Trump signs an executive order establishing voluntary 30-day pre-release review of frontier models for national security. Gemini 3.5 Flash goes generally available in GitHub Copilot, escalating the developer-tooling platform war.
Read →AI Briefing: June 2, 2026 — The First AI Agent Attack, Apple's Gemini Bet, and Nvidia Beyond the GPU
Sysdig documents the first autonomous LLM-agent-driven cyberattack in the wild: four pivots, one database exfiltrated, under two minutes, no human in the loop. Apple prepares to unveil a Gemini-powered Siri at WWDC on June 8 — the most consequential AI alliance in the consumer platform market. Nvidia's Vera CPUs enter full production with Anthropic and OpenAI as customers, completing the company's push from GPU vendor to full-stack AI silicon provider.
Read →AI Weekly: May 26–June 1, 2026 — The Bill Arrives, the Enterprise Splits, and OpenAI Writes Its Own Rules
GitHub Copilot's metered billing goes live today, ending unlimited AI coding. KPMG embeds Claude in 276,000 employees' workflows while Microsoft cancels its Claude Code licences over $2,000/month bills. OpenAI publishes its Frontier Governance Framework ahead of the IPO. Altman admits he was wrong on jobs. DeployCo begins its first client engagements.
Read →AI Weekly: May 25–31, 2026 — Anthropic Hits $965 Billion, Karpathy Crosses the Aisle, China Locks Its Researchers In, and Google Retires the Search Box
Anthropic's Series H closes at $965B — the largest venture round in history — overtaking OpenAI's private valuation. Karpathy joins Anthropic's pretraining team. China formalises travel restrictions on private-sector AI researchers. Google routes global search through Gemini 3.5 Flash, retiring the traditional search box. OpenAI turns on advertising inside ChatGPT.
Read →AI Weekly: May 26–30, 2026 — Opus 4.8 Lands, Spark Goes Live, and the Labs Walk Back the Apocalypse
Anthropic ships Claude Opus 4.8 with dynamic workflows and a honesty upgrade that makes it four times less likely to miss its own bugs. Gemini Spark goes live for US AI Ultra subscribers. Altman and Amodei reverse on the AI jobs apocalypse the same week OpenAI files its S-1. OpenAI's unit economics face public scrutiny. Meta's $135B infrastructure commitment: the week in five stories.
Read →AI Briefing: May 27, 2026 — Altman's Jobs Reversal, Google's EU Reckoning, and Meta's $135B Infrastructure Bet
Sam Altman tells Sydney he was wrong about the AI jobs apocalypse — and explains why in ways that directly contradict Suleyman's 18-month automation clock. Brussels finalises a nine-figure Google fine with structural AI search remedies. Meta commits $135B in capex to close the AI gap. Three stories, one accelerating industry.
Read →AI Briefing: May 26, 2026 — Magnifica Humanitas: What the Church's AI Manifesto Actually Demands
Pope Leo XIV's 42,300-word encyclical on AI is now public, co-presented with Anthropic's Chris Olah at the Vatican. Here is what the document actually says on regulation, warfare, and workers — why Anthropic chose this stage over the White House — and what it means when the world's oldest institution enters the AI governance arena with institutional weight.
Read →AI Weekly: May 19–25, 2026 — The Church Speaks, Two Labs Race for $1 Trillion, and the Grid Bets on AI
Pope Leo XIV publishes Magnifica Humanitas, the first AI ethics encyclical, with Anthropic co-founder at the Vatican. Google I/O delivers Gemini Omni and Antigravity 2.0. OpenAI files its S-1 at over $1 trillion. Anthropic approaches $900B valuation. NextEra acquires Dominion for $67B to power AI data centres. The week in five stories.
Read →AI Weekly: May 19–24, 2026 — Washington Retreats, NVIDIA Soars, and the Industry Sets Its Own Terms
Trump kills the AI executive order hours before signing. NVIDIA posts $81.6B in quarterly revenue, up 85% year-on-year. Anthropic and the Gates Foundation commit $200M to AI for global development. Gemini Spark gains MCP support for third-party apps. A supply chain attack exfiltrates 3,800 GitHub repos in eighteen minutes. The week in five stories.
Read →AI Weekly: May 18–23, 2026 — Google Rewires the Platform, Musk Loses in Court, and OpenAI Heads for a Trillion-Dollar IPO
Google I/O delivers Gemini Omni and Antigravity 2.0, redefining the platform. A jury dismisses Musk's OpenAI lawsuit in under two hours. Karpathy joins Anthropic's pretraining team. Meta cuts 8,000 jobs and bets $145B on AI infrastructure. OpenAI files its S-1. The week in five stories.
Read →AI Briefing: May 22, 2026 — OpenAI's Trillion-Dollar IPO: What the S-1 Has to Explain
OpenAI files its S-1 confidentially with the SEC, targeting a September 2026 listing at over $1 trillion. $25B in annualised revenue, $14B in annual losses, Goldman Sachs and Morgan Stanley advising. The most consequential technology IPO in years — and the questions the prospectus cannot dodge on unit economics, competitive position, and a governance structure that has no precedent.
Read →AI Briefing: May 20, 2026 — Pope Leo XIV's AI Doctrine, Anthropic's Revenue Supernova, and the First Models to Clear the Cyberattack Gauntlet
Pope Leo XIV publishes Magnifica Humanitas on May 25 — the Church's first AI ethics encyclical — with Anthropic co-founder Christopher Olah at the Vatican. Anthropic hits $30B ARR doubling every six weeks, reportedly passing OpenAI in revenue. Claude Mythos becomes the first AI model to clear a 32-step corporate network attack simulation. Three institutions, one threshold.
Read →AI Briefing: May 19, 2026 — Google I/O's Intelligence Gambit, the Googlebook, and OpenAI at $852B
Google I/O 2026 keynote reframes Android 17 as an intelligence system with Gemini running beneath every app. Google enters the premium AI PC market with Googlebook and previews Android XR glasses. OpenAI plots a $1 trillion IPO while losing $14 billion a year. Three stories, one platform war.
Read →AI Briefing: May 18, 2026 — Suleyman's 18-Month Clock, JPMorgan's $20B Bet, and the End of OpenAI-Microsoft Exclusivity
Microsoft's AI chief sets an 18-month deadline for white-collar automation. JPMorgan reclassifies AI as core infrastructure alongside cybersecurity, committing $19.8B in 2026 with 500+ use cases in production. OpenAI formally ends its Microsoft exclusivity to sell on AWS and Google Cloud. Three stories, one direction.
Read →AI Weekly: May 11–17, 2026 — Trillion-Dollar Bets, Android Reborn, and the First AI Zero-Day
Anthropic closes in on a $950B valuation. Google I/O declares Android an intelligence system built around Gemini. Researchers confirm the first AI-assisted zero-day in the wild. OpenAI ships GPT-5.5. The Pentagon opens classified networks to seven AI companies. The week in five stories.
Read →AI Briefing: May 16, 2026 — Anthropic's Near-Trillion Round, Google I/O, and GPT-5.5
Anthropic enters talks for a $950B valuation that would surpass OpenAI, commits $200M to the Gates Foundation for global health and education AI, Google I/O declares Android an intelligence system built around Gemini, and OpenAI ships GPT-5.5. Four stories, one accelerating industry.
Read →AI Briefing: May 14, 2026 — Mythos, Android Gemini, and the First AI Zero-Day
Anthropic's Claude Mythos Preview visits the White House and anchors a $1.5B financial services JV. Google quietly rebuilds Android around Gemini. Researchers confirm the first AI-assisted zero-day in the wild. Novo Nordisk puts its entire drug pipeline on OpenAI. Five stories, one direction.
Read →The GM IT Skills Swap: A Blueprint for How AI Transforms Corporate Tech Teams
General Motors cut 600 IT workers and immediately started hiring AI-native engineers, data specialists, and agent developers. The pivot to software-defined vehicles built on Google Gemini and Nvidia Drive Thor is driving the restructure — and the template is already spreading.
Read →OpenAI Daybreak: AI That Hunts Vulnerabilities in Your Code
OpenAI launched Daybreak — a cyber defence platform on GPT-5.5 and Codex Security that finds, validates, and proposes patches for vulnerabilities across entire codebases. Hours of analysis to minutes. Partners include CrowdStrike, Palo Alto, Cloudflare, and Cisco. The dual-use question comes standard.
Read →The OpenAI Deployment Company: When a Lab Becomes a Consulting Firm
OpenAI raised $4B at a $14B valuation to create a dedicated enterprise deployment arm, acquired AI consulting firm Tomoro, and brought in McKinsey, Bain, Goldman Sachs, and SoftBank as investors. The lab is becoming the integrator — and the implications for everyone else in enterprise AI are significant.
Read →The AI That Outdiagnosed the Doctors: What the Harvard ER Study Actually Proved
A Harvard study in Science tested OpenAI's o1 against real emergency room cases. The AI scored 67% on correct diagnoses. Physicians scored 50–55%. The triage advantage was the most surprising finding — and the limitations are the most important part of the story.
Read →When Your AI Agent Is the Attack Surface: The Five Eyes Guidance on Agentic AI
CISA, NSA, and their Five Eyes counterparts just published the first joint security guidance on agentic AI. Prompt injection, over-privileged agents, supply chain risk — they're naming them because they're already observing the consequences.
Read →The Stanford AI Index 2026: What the Numbers Actually Say
SWE-bench near 100%. $581B invested. Frontier models winning gold at the IMO. But transparency scores dropped from 58 to 40, and the same model that aced the Olympiad reads analog clocks correctly only 50.1% of the time. Stanford HAI's annual report, distilled.
Read →The Pentagon AI Deals: What Happens When Safety Is a Dealbreaker
The DoD signed classified AI contracts with OpenAI, Google, Microsoft, Nvidia, AWS, Oracle, SpaceX, and Reflection — and froze out Anthropic for refusing to drop safety guardrails on autonomous weapons. It labeled Anthropic a "supply chain risk." A federal court blocked it. The dispute isn't over.
Read →The Goblin in the Machine: What ChatGPT's Creature Fixation Reveals About AI Training
ChatGPT spent months inserting goblins, gremlins, and trolls into unrelated conversations. OpenAI's post-mortem traces the cause to a Nerdy personality training signal that leaked through RLHF into the base model — a perfect real-world example of reward hacking.
Read →The Context Window War: What Million-Token Models Actually Change
Claude hits 1M tokens. Gemini stretches to 2M. Raw context capacity matters — but the retrieval-versus-stuffing decision, the lost-in-the-middle problem, and the cost realities are more nuanced than the headline numbers.
Read →AI Coding Agents in 2026: From Autocomplete to Autonomous Pull Requests
Today's coding agents open PRs, write tests, and debug CI pipelines. Here's an honest look at what the current generation — Claude Code, Copilot Workspace, Cursor — actually does well, and where human review remains irreplaceable.
Read →EU AI Act: What Product Teams Need to Do Before August 2026
GPAI transparency obligations are already in force. High-risk enforcement begins August 2026. Here's how to read the risk classification framework, what conformity assessment actually requires, and a 90-day action checklist.
Read →Llama 4: The Open-Weights Comeback
Llama 4 Scout and Maverick use a Mixture of Experts architecture that competes with proprietary APIs on major benchmarks — at open-weights cost. Here's what the numbers show, what the licence actually says, and what it changes for independent teams.
Read →Claude Opus 4.7: What Adaptive Thinking Changes About AI Product Design
Opus 4.7 removed the fixed thinking budget model entirely. Adaptive thinking, the new effort parameter with its xhigh tier, hidden thinking content by default, and task budgets — here's what changes about how you architect AI features.
Read →Salesforce Headless 360
Lightning Experience is powerful but it's not always the right UI for the job. Here's how we architect Headless 360 — using Salesforce as the data and logic layer while owning the frontend entirely.
Read →Using the Claude SDK: A Practical Guide
The Anthropic SDK is clean, well-designed, and increasingly our default for AI-powered features. Streaming, tool use, prompt caching, vision — and the patterns that actually hold up in production.
Read →Using the OpenAI SDK: A Practical Guide
The OpenAI SDK is the most-used AI SDK in production today. Chat completions, function calling, Structured Outputs, the Assistants API — and the cost management lessons we learned the hard way.
Read →Building GitHub Copilot Extensions: A Practical Guide
Copilot Extensions let you build @agents that appear natively inside Copilot Chat. The request format, SSE streaming protocol, context injection, and what the official docs quietly skip over.
Read →AI in 2025: The Year Reasoning Won
Reasoning models matured, vibe coding hit a wall, open weights closed the quality gap, and AI tooling became infrastructure. Our honest accounting of the year that changed everything.
Read →Multi-Agent Systems in Production: What Actually Works
The vision of AI agent swarms ran ahead of the engineering reality. Here's what multi-agent architectures look like when they work — and what consistently fails.
Read →OpenAI o3: The Reasoning Model That Aced ARC-AGI
o3 scored 87.5% on ARC-AGI, a benchmark designed to resist AI. Here's what that score actually means, what o3 costs to use, and what it means for the applications you're building.
Read →DeepSeek V3: The Model That Changes the Economics of AI
Trained for $5.6M on export-restricted hardware, matching GPT-4o on major benchmarks. DeepSeek V3 isn't just a model release — it's a challenge to the capital-intensity thesis of frontier AI.
Read →Copilot Workspace vs Cursor: The IDE Wars Go Agentic
GitHub's Copilot Workspace went GA in September. Cursor has been in market for two years. They're targeting different moments in the workflow — here's how to think about the choice.
Read →Salesforce Agentforce: What Einstein Copilot's Rebrand Actually Means
Salesforce launched Agentforce at Dreamforce with autonomous agent capabilities, Atlas reasoning, and $2-per-conversation pricing. Here's what changed, what didn't, and what it means for builders.
Read →Building With MCP: Real-World Agent Integration
MCP went from Anthropic's proposal to an industry standard in months. By September 2025 there were 2,000+ published servers. Here's what building and consuming them actually looks like in production.
Read →Gemini 2.5 Pro: Google's Most Capable Model Yet
63.8% on SWE-Bench Verified, a 2M token context window, and genuine multimodal reasoning. Google's best model has closed most of the gap. Here's where it leads and where it still lags.
Read →Claude's 1M Token Context Window: Promise vs. Practice
A million tokens is more than War and Peace. The capability is real and the prompt caching economics are compelling. The attention distribution tradeoffs and cost realities are more nuanced than the headline.
Read →CrewAI, LangChain, AutoGPT: Which Agent Framework Should You Use?
CrewAI raised $18M with 280% growth. LangChain pivoted to LangGraph. AutoGPT found direction. The landscape has consolidated — here's how to choose, and when to skip the framework entirely.
Read →The Vibe Coding Hangover
A METR study found AI tools made experienced developers 19% slower on realistic tasks. The vibe coding wave was real. So was the reckoning. Here's what the research actually shows.
Read →Cursor: The AI-First Editor Eating VS Code's Lunch
Cursor crossed 1 million users by treating AI as the primary interface, not the add-on. Multi-file edits, codebase indexing, agent mode — and when it's the better choice over VS Code plus Copilot.
Read →Running Local LLMs With Ollama: A Production Guide
Ollama made running local models trivially easy. Production is harder — concurrency, versioning, monitoring, model selection, and the hybrid routing pattern that gets you the best of both worlds.
Read →RAG in 2025: From Retrieval to Context Engines
Naive vector search worked well enough in 2023. By mid-2025, the teams winning with RAG had moved far beyond cosine similarity — hybrid retrieval, re-ranking, agentic loops, and contextual compression.
Read →pgvector vs Pinecone: Choosing a Vector Store for RAG
Most teams reach for a dedicated vector database too early. Here's the decision framework: when pgvector inside Postgres is genuinely enough, and when you actually need Pinecone, Weaviate, or Qdrant.
Read →How Reasoning Models Work: Chain of Thought at Scale
From o1's 74% AIME to o3's 96% and Gemini 2.5 Pro's 63.8% SWE-Bench. The chain-of-thought revolution explained — training, benchmarks, cost tradeoffs, and the 2025 landscape.
Read →GitHub Copilot Agent Mode: From Autocomplete to Autonomous Developer
Announced in February 2025, Copilot Agent Mode gives the model a terminal and a multi-file edit loop. Here's what changed, what it means in practice, and where it still needs a human in the loop.
Read →Model Context Protocol: Anthropic's Standard for AI Tool Integration
Announced November 2024, adopted by OpenAI and Google by April 2025, donated to the Linux Foundation by December. How MCP became the USB-C of AI tool integration in under a year.
Read →How We Built the Salesforce Screen Recorder
A look inside the architecture of our most-used Chrome extension — from capturing Lightning page state to bundling structured ZIPs that GitHub Copilot can actually reason about.
Read →When to Use AI Agents (and When Not To)
Agents are powerful and overhyped in equal measure. Here's the four-question filter we apply before reaching for an agentic architecture on any client project.
Read →Vibe Coding: Writing Software by Describing It
Coined by Karpathy in February, named Collins' Word of the Year by November. 25% of YC W25 startups had 95%+ AI-generated codebases. Here's what vibe coding actually is and what it's not.
Read →DeepSeek R1: The Open-Source Reasoning Breakthrough
Released January 21, matched OpenAI o1 on AIME (79.8% vs 79.2%), available under MIT licence, API at $0.55/$2.19 per 1M tokens vs o1's $15/$60. The model that changed the reasoning model landscape.
Read →Offline-First Mobile: What Nobody Tells You
Building Tasted taught us a lot about offline-first architecture on React Native. Conflict resolution, sync strategies, and why "no account required" is a product decision as much as a technical one.
Read →Generating Salesforce Docs From Metadata
How we built SF Docs Portal — a static site generator that turns raw Salesforce repository metadata into a fully searchable architecture documentation site. Zero backend, zero config.
Read →Boutique Studio vs. Agency: Why We Chose Small
We could have grown into an agency. We chose not to. The case for staying small, staying sharp, and why "boutique" isn't a polite word for "we can't scale".
Read →