エピソード

  • Ep. 95 — Gemini 3.5 goes agentic, the $900B memory crunch, and half-price Claude
    2026/06/30
    Gemini 3.5 goes agentic, the $900B memory crunch, and half-price Claude Daily Prompt with Archer & Iris — Tuesday, June 30, 2026 (Daily Briefing) Google shipped a new frontier model that doesn't just answer — it acts. But the louder story is everyone else fighting over the rest of the stack: the chips underneath it, the governments on top of it, and the models in the middle. The thread today is ownership. LEAD STORIES - Google ships Gemini 3.5, pitched as "intelligence with action." You hand it a goal; it plans the steps and spins up parallel subagents under a supervisor — maintaining a codebase, prepping financial documents, sorting assets. Google says it beats its own last flagship on coding and agentic tasks and outputs about 4x faster than other frontier models. Free in the Gemini app and Search, and on the developer API day one. The catch builders should sit with: when the model takes the action, a mistake is no longer a wrong sentence — it's a wrong action, already taken. https://deepmind.google/blog/gemini-3-5-frontier-intelligence-with-action/ - South Korea commits $900B+ to ease "RAMageddon" — an AI-driven global memory shortage. The plan includes about $518B for four new memory fabs (plus ~$52B for an HBM packaging hub and ~$356B for AI data centers through 2035). President Lee called semiconductors, physical AI, and data centers the "triple axis," noting existing plants have "already reached their limits." The market already moved: memory maker Micron's stock jumped 236% in a month, revenue quadrupled to ~$41.45B in a quarter, and its market cap briefly topped Meta and Tesla — earning the "next Nvidia" label. The shortage runs into 2027. The bottleneck moved from the chip that thinks to the memory that feeds it — and it's spilling into laptop and phone prices. https://techcrunch.com/2026/06/29/south-korean-tech-giants-commit-over-550b-to-ease-ramageddon/ https://techcrunch.com/2026/06/28/why-wall-street-thinks-us-memory-maker-micron-is-the-next-nvidia/ - Anthropic and Gov. Newsom give California state agencies and local governments Claude at half price. Newsom: AI "should not replace the human work of government; it should help our workers move faster... and deliver better results for Californians." The tell: Anthropic is the vendor the federal government has handled more cautiously on supply-chain risk, yet California's CIO said that designation "just didn't come up" — while the Pentagon's marquee AI deal went to OpenAI. Same company, two governments, opposite risk readings — and states are the next anchor customer the labs are racing to sign. https://techcrunch.com/2026/06/29/anthropic-and-gov-newsom-forge-deal-allowing-california-government-to-use-claude-at-half-price/ QUICK HITS - Vibe-coding platform Base44 (acquired by Wix for $80M last year) launched its own model, Base1, trained on tens of millions of user sessions — aiming to be cheaper and faster than renting frontier models. Base44 is at $100M run-rate revenue; rival Lovable is at $500M. https://techcrunch.com/2026/06/29/vibe-coding-platform-base44-launches-own-model-as-ai-startups-seek-defensibility/ - Microsoft launched MAI-Code-1-Flash, a lightweight agentic coding model native to VS Code and Copilot — another first-party model that does OpenAI's job, hedging Microsoft's biggest partnership. https://microsoft.ai/models/mai-code-1-flash/ - Crypto exchange OKX opened OKX AI, a marketplace where AI agents hire and pay each other in stablecoins around the clock, with on-chain reputations. It launched to developers after a 50-provider beta; the CEO is betting on million-dollar one-person companies. https://techcrunch.com/2026/06/30/crypto-exchange-okx-wants-ai-agents-to-hire-and-pay-each-other/ - Arena, the AI leaderboard everyone uses, is now a $100M run-rate business — eight months after it started charging, on 10M+ human votes. The CEO's caveat: the revenue "is not recurring," and most people still think it's a free open-source project. https://techcrunch.com/2026/06/29/arena-the-ai-leaderboard-everyone-uses-is-now-a-100m-business/
    続きを読む 一部表示
    7 分
  • Ep. 94 — The frontier behind a federal gate — and chatbot logs go to trial
    2026/06/29
    The frontier behind a federal gate — and chatbot logs go to trial Daily Prompt with Archer & Iris — Monday, June 29, 2026 (Daily Extended) Two of the most powerful AI models shipped this weekend, and both went out behind a government-approved guest list. The other half of the weekend played out in a Los Angeles courtroom, where a defendant's chatbot history was entered as evidence. The thread: permission — who's cleared to run the best models, and whose private chats can be read back to them. LEAD STORIES - OpenAI unveils GPT-5.6 — Sol, Terra, and Luna — but ships it as a government-cleared limited preview. Sol is pitched at stronger coding, science, and cybersecurity, with a new subagent "ultra" mode; OpenAI claims it edges Anthropic's best on code using about a third fewer output tokens. The launch went only to trusted partners whose names were shared with Washington first. OpenAI: a government access process like this shouldn't become "the long-term default." https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/ https://www.theverge.com/ai-artificial-intelligence/957845/openai-gpt-5-6-trump-administration-ai-preview - The Trump administration clears Anthropic's Mythos 5 — its strongest cybersecurity model — for more than 100 U.S. companies and agencies, including their non-American staff, two weeks after the Mythos-class ban dated June 12. The public-facing Fable 5 remains offline (15 days, expected back this week). Meanwhile Tokyo's Sakana AI (Fugu) and a Chinese security firm shipped Mythos-like tools — "frontier capability without the export-control risk." Anthropic's run-rate revenue hit $47 billion as of May 2026. https://techcrunch.com/2026/06/26/trump-admin-releases-anthropic-mythos-to-be-used-by-more-than-100-us-companies-agencies/ https://techcrunch.com/2026/06/27/asian-ai-startups-launch-mythos-like-models-as-anthropics-export-ban-drags-on/ - Prosecutors used a defendant's ChatGPT logs as evidence in the Palisades fire arson trial. The Palisades Fire was one of LA's deadliest — about 12 killed, 6,000+ structures lost. Investigators reviewed thousands of conversations he used like a diary (fire images; "Why am I so angry all the time?"). The jury deadlocked 10–2 to acquit; a mistrial was declared, with a retrial set for October 19, 2026. https://www.theverge.com/ai-artificial-intelligence/958751/prosecutors-chatgpt-palisades-wildfire-arson-mistrial QUICK HITS - OpenAI is leaning toward pushing its IPO to 2027 as tech-stock volatility rattles advisers — a ~$1T target on ~$2B/month revenue, last valued near $852B in March 2026. https://www.pymnts.com/news/investment-tracker/ipo/2026/openai-weighs-delay-of-ipo-as-tech-stock-volatility-rattles-advisers/ - Apple's Vision Pro chief (Paul Meade), who oversaw the headset and Apple's coming smart glasses, is reportedly leaving for OpenAI's hardware team after a hardware-org reshuffle. https://techcrunch.com/2026/06/27/apple-vision-pro-exec-is-reportedly-leaving-for-openai/ - China's Z.ai released open-weight GLM-5.2; researchers say it matches Anthropic on certain cybersecurity tasks, though it lags on general capability — the exact capability the U.S. is restricting, given away free. https://www.theverge.com/ai-artificial-intelligence/958804/chinas-z-ai-glm-52-mythos-cybersecurity - Suno launched Spark, an incubator offering grants, mentorship, and marketing to unsigned artists, as it pushes to become a label and streaming destination. https://www.theverge.com/ai-artificial-intelligence/958801/suno-launches-spark-incubator-program-to-feed-independent-artists-to-its-ai-machine - HP scaled its OpenAI "Frontier" partnership, wiring the models across customer support, software development, and enterprise operations. https://openai.com/index/hp-frontier-partnership
    続きを読む 一部表示
    8 分
  • Ep. 93 — Deep Dive: Who Owns the Simulated World (Part 2 — The Money and the Moat)
    2026/06/28
    Deep Dive: Who Owns the Simulated World (Part 2 — The Money and the Moat) Daily Prompt with Archer & Iris — 2026-06-28 Part 2 of a weekend pair on world models. Saturday asked who's building AI that predicts the next state of the world instead of the next word. Today follows the money: more than three billion dollars poured into the category in 2026 — most of it before anyone shipped a product you can buy — and the quiet fight over who actually captures the value when world generation becomes an API call. In this episode: - The money: Yann LeCun's Paris lab, AMI Labs, closed a $1.03 billion seed at a $3.5 billion pre-money valuation — the largest seed round in European history — with no product, on the strength of his thesis. Backers include Cathay Innovation, Greycroft, Hiro Capital, HV Capital, Bezos Expeditions, NVIDIA, Temasek, Samsung and Toyota Ventures, plus Jeff Bezos, Eric Schmidt and Tim Berners-Lee. Fei-Fei Li's World Labs has raised ~$1.23B and was reported in talks near a $5B valuation; Odyssey raised $310M at $1.45B; Decart raised $300M at ~$4B. The motive read: NVIDIA sits on the AMI, World Labs and Decart cap tables — funding its own future customers. https://techcrunch.com/2026/03/09/yann-lecuns-ami-labs-raises-1-03-billion-to-build-world-models/ https://news.crunchbase.com/venture/world-model-ai-lab-ami-raises-europes-largest-seed-round/ - The moat: NVIDIA launched Cosmos 3 (May 31, 2026), "the world's first fully open omnimodel," and posted it free on Hugging Face — then convened the Cosmos Coalition (Black Forest Labs, Runway, LTX, Generalist, Agile Robots, Skild AI). The reframe: NVIDIA commoditizes the model layer it doesn't need to own to defend the chip layer it does. Jensen Huang: "The big bang of physical AI is just around the corner." The counter-move: Odyssey took Amazon's money, made AWS its preferred cloud, and is optimizing for Amazon's own silicon to escape NVIDIA hardware. https://nvidianews.nvidia.com/news/nvidia-launches-cosmos-3-the-open-frontier-foundation-model-for-physical-ai https://techcrunch.com/2026/06/17/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names/ - The commoditization and the builder's move: Decart (~$4B valuation) opened its Oasis 3 driving world model via API at two cents per second, aiming to build a developer ecosystem "much like how OpenAI did with language models" — its earlier model, Lucy, already has 100,000+ developers. The caveats are real: scenery degrades, control drops, and "the car will just drive through other cars." World-model and inference startups pulled ~$1.8B in two days in June as foundation models commoditize. World Labs sells Marble from $20–$95/month and anchored its $1B Series B with $200M from Autodesk — a path into existing CAD/design customers. Builder takeaway: don't train your own; pick the layer with lasting memory, a real SDK, a stable price, and distribution. https://techcrunch.com/2026/06/10/decarts-new-world-model-can-simulate-hours-of-photorealistic-driving-with-some-caveats/ https://www.techtimes.com/articles/318779/20260621/ai-inference-world-model-startups-pull-18b-two-days-foundation-models-commoditize.htm https://siliconangle.com/2026/01/23/fei-fei-lis-world-labs-reportedly-raising-funding-5b-valuation/
    続きを読む 一部表示
    7 分
  • Ep. 92 — Deep Dive: When AI Stopped Predicting Words (Part 1 — The World Model Race)
    2026/06/27
    Deep Dive: When AI Stopped Predicting Words (Part 1 — The World Model Race) Daily Prompt with Archer & Iris — 2026-06-27 For three years the smartest people in AI argued about what the next word should be. This year a lot of them quietly stopped asking. Part 1 of a weekend pair on world models — AI that predicts the next state of the world instead of the next word, trading autocomplete-for-sentences for autocomplete-for-outcomes. Today: where the idea came from, why language hit a wall, and who's building the alternative. Tomorrow (Part 2): the money, the moat, and who walks away owning the simulated world. In this episode: - The category shift — world models vs. LLMs: a language model predicts the next word and has no grounded model of physics; a world model learns from video and predicts how a scene changes when you act on it, unlocking planning, physics, and cause-and-effect. Yann LeCun: "We are not going to get to human-level AI just by scaling LLMs. Silicon Valley is completely hypnotized." His JEPA approach predicts in an abstract latent space, not pixels. Also driving the timing: text is hitting a "peak data" ceiling while video is barely tapped. https://introl.com/blog/world-models-race-agi-2026 - The cast (incumbents): DeepMind's Genie generates walkable, photorealistic 3D worlds from a prompt in real time at 24 fps — object permanence holds for roughly a minute before drift; Project Genie opened to Google AI Ultra subscribers in the US in January 2026. NVIDIA's Cosmos (launched CES 2025) is the training-ground/synthetic-data backbone — ~20 million hours of video, 2M+ downloads by January 2026 — powering humanoid players like Figure and 1X. https://www.ai.cc/blogs/world-models-2026-google-nvidia-physical-ai-breakthroughs/ - The cast (challengers): Fei-Fei Li's World Labs shipped Marble (Nov 2025) — generate a brand-new 3D world from text/image/sketch and export it to Unreal, Unity, Vision Pro, or Quest 3, on a free + paid pricing tiers — her bet on "spatial intelligence, the frontier beyond language." Odyssey shipped a commercial world-model API (JS/Python SDKs) sustaining 120+ seconds of coherent simulation. More than $1.3B flowed into world-model startups in early 2026. https://time.com/7339513/ai-fei-fei-li-virtual-worlds/ - Why it matters for builders — Yann LeCun left Meta after 12 years to start AMI Labs (Paris; CEO Alex LeBrun), raising ~€500M at a ~€3B valuation before shipping a product. The concrete payoff: Figure trains its robots end-to-end with reinforcement learning in simulation, then transfers the policy to the physical robot with no real-world fine-tuning — closing the "sim-to-real gap." The reframe: the warehouse stops being the classroom and becomes the final exam. The catch: predicting a world costs far more compute than predicting a sentence. https://ai2.work/technology/world-models-in-2026-why-lecun-fei-fei-li-and-deepmind-bet-billions-on-3d-ai/ The builder question for Part 2 (tomorrow): when world generation becomes an API call, who captures the value — the lab that trains the model, the cloud that rents the GPUs, or the builder who finds the use?
    続きを読む 一部表示
    8 分
  • Ep. 91 — Washington paces GPT-5.6, Anthropic vs. Alibaba, and a 1,000x power claim
    2026/06/26
    Daily Prompt with Archer & Iris Friday, June 26, 2026 — Daily Briefing Washington tells OpenAI to pace GPT-5.6, Anthropic accuses Alibaba of the biggest Claude copy job yet, and a Databricks alum says he can cut AI's power bill a thousandfold. LEAD STORIES - White House asks OpenAI to slow-roll GPT-5.6 — a customer-by-customer limited preview over cybersecurity concerns about models finding and exploiting software vulnerabilities. https://techcrunch.com/2026/06/25/the-white-house-is-asking-openai-to-slow-roll-the-release-of-its-new-model-over-safety-concerns/ - Anthropic tells the Senate that Alibaba's Qwen lab ran the largest Claude distillation attack to date — ~25,000 fake accounts, 28.8M exchanges (Apr 22–Jun 5), targeting agentic reasoning and coding. https://arstechnica.com/tech-policy/2026/06/anthropic-claims-alibaba-defied-trump-to-attack-claude-and-steal-capabilities/ - Naveen Rao's Unconventional AI claims a 1,000x cut to AI inference power via oscillator-based compute; first demo (Un-0 image model) runs in simulation, no chip yet. https://techcrunch.com/2026/06/25/databricks-former-ai-chief-thinks-he-can-cut-ais-power-bill-by-1000x/ QUICK HITS - Claude is winning paid consumers: paid revenue up ~75% since January; Claude course demand up 18x in a month and now outsearching "AI." https://techcrunch.com/2026/06/25/anthropics-claude-is-winning-over-paid-consumers-a-market-owned-by-chatgpt/ - Adobe acquires Topaz Labs (Astra, Wonder, on-device NeuroStream); tools fold into Firefly, deal closes H2 2026. https://techcrunch.com/2026/06/25/adobe-acquires-image-and-video-enhancement-tool-maker-topaz-labs/ - Ford takes No. 1 in JD Power initial quality — and credits rehiring engineers to fix mistakes from its automated systems. https://www.theverge.com/transportation/956316/ford-quality-jd-power-ranking-ai-automated-mistakes - IBM claims the world's first sub-1-nanometer chip technology using stacked nanosheet transistors. https://arstechnica.com/gadgets/2026/06/ibm-claims-worlds-first-sub-1-nanometer-chip-technology/ - General Intuition raises $320M at a $2.3B valuation to train robots on action-labeled video-game footage; backers include Bezos and Eric Schmidt. https://techcrunch.com/2026/06/25/general-intuitions-2-3b-bet-that-video-games-can-train-ai-agents-for-the-real-world/
    続きを読む 一部表示
    8 分
  • Ep. 90 — OpenAI builds a chip, Google's cheap agent, and the AI cost bill comes due
    2026/06/25
    Daily Prompt with Archer & Iris — 2026-06-25 (Thursday / Daily Briefing) The whole day points one direction: the cost of running these models. OpenAI builds its own chip, Google pushes computer-use into its cheapest tier, and the receipts start arriving everywhere else. In this episode: LEADS - OpenAI unveils Jalapeño, its first custom inference chip, built with Broadcom. Inference-only, claims much better performance-per-watt, designed with help from OpenAI's own models — and one more step toward owning the whole stack. https://openai.com/index/openai-broadcom-jalapeno-inference-chip - Google brings computer use to Gemini 3.5 Flash. Agentic screen control — click, type, navigate across browser, mobile and desktop — lands in the fast, cheap tier, with prompt-injection halting and confirmation on irreversible actions. On the Gemini API now. https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-computer-use-gemini-3-5-flash/ - Engineering is the most AI-resilient job, per SignalFire's State of Talent 2026. Engineering hiring down only 11% vs 2019 (all tech down 25%); engineers now 55% of new hires at the majors. Jevons paradox, with the squeeze landing on mid-stage and non-engineering roles. https://techcrunch.com/2026/06/24/ai-was-supposed-to-kill-engineering-jobs-but-new-data-suggests-theyre-the-most-resilient/ QUICK HITS - Cerebras stock falls ~20% on its first earnings as margin guidance is cut to the high thirties; revenue still up 94% to $193M. CEO says investors misunderstood. https://techcrunch.com/2026/06/24/cerebras-stock-plunges-after-earnings-as-ceo-says-margin-outlook-was-misunderstood/ - From tokenmaxxing to token rationing: Accenture restricts AI tokens, Amazon deletes its usage leaderboard, as spend becomes material to the cost structure. https://techcrunch.com/2026/06/24/companies-are-scrambling-to-stop-employees-from-maxing-out-ai-budgets-with-small-tasks/ - Meta pauses its Model Capability Initiative — keystroke and mouse tracking for AI training data — after an internal leak exposed private conversations and performance data company-wide. https://www.engadget.com/2199458/meta-is-pausing-employee-tracking-program-after-it-let-the-whole-company-see-sensitive-data/ - Figma adds code layers, animation and AI shader tools to the design canvas, thinning the design-to-code handoff further. https://techcrunch.com/2026/06/24/figma-adds-code-layers-support-for-animations-more-ai-features-in-new-update/
    続きを読む 一部表示
    8 分
  • Ep. 89 — Claude Tag moves into Slack, AI's $27M primary, and the Altman film no one will touch
    2026/06/24
    Daily Prompt with Archer & Iris Ep. 89 — Claude Tag moves into Slack, AI's $27M primary, and the Altman film no one will touch Wednesday, June 24, 2026 Today's stories — three angles on the same question: who's holding the leash. LEADS - Anthropic launches Claude Tag, an ambient agent that lives in your Slack channels, learns your company over time, and runs on a new shared, permissioned "agent identity" model — the layer enterprises were waiting for before trusting agents. https://techcrunch.com/2026/06/23/anthropics-claude-tag-is-learning-your-company-one-slack-message-at-a-time/ https://claude.com/blog/agent-identity-access-model - AI super PACs spent about $27 million on a single NYC House primary (NY-12) to defeat RAISE Act author Alex Bores — and he lost. A deterrent priced for all 535 members of Congress. https://www.theverge.com/policy/954970/ai-super-pacs-alex-bores-new-york-12th-district - Hollywood passes on "Artificial," the $40M Luca Guadagnino film with Andrew Garfield as Sam Altman. Amazon, Netflix, A24, Focus, and Warner all stepped away — months after Amazon's $50B OpenAI investment, with A24 backed by a firm holding an OpenAI board seat. https://www.theverge.com/entertainment/954899/luca-guadagnino-artificial-sam-altman-amazon-a24-neon-mubi-chatgpt QUICK HITS - Amazon runs its first-ever ads inside ChatGPT — the store and the assistant become one screen. https://www.businessinsider.com/amazon-first-chatgpt-ads-ai-strategy-openai-data-2026-6 - Meta launches cheaper smart glasses, dropping the Ray-Ban name to push AI eyewear mainstream. https://www.theverge.com/tech/954052/meta-glasses-hands-on-kylie-jenner-smart-glasses-price-battery-privacy - Fika raises $4M for a hiring platform where AI agents conduct the interviews. https://techcrunch.com/2026/06/23/fika-jobs-raises-4m-to-build-a-video-first-hiring-platform-where-ai-agents-interview-candidates/ - SoftBank's Masayoshi Son questions Elon Musk's plan for AI data centers in space — the doubt now coming from the believers. https://www.businessinsider.com/softbank-ceo-masayoshi-son-elon-musk-ai-space-data-centers-2026-6 Daily Prompt with Archer & Iris — the best AI news for builders like you.
    続きを読む 一部表示
    7 分
  • Ep. 88 — OpenAI patches open source, Nvidia hollows out Groq, SpaceX rents to a rival
    2026/06/23
    Ep. 88 — OpenAI patches open source, Nvidia hollows out Groq, SpaceX rents to a rival Daily Prompt with Archer & Iris — 2026-06-23 (Tuesday / Daily Briefing) In this episode: Leads - OpenAI launches Daybreak and its open-source arm "Patch the Planet" — partnering with security firm Trail of Bits to scan open-source code, triage the findings, and write fixes with maintainers. It lands directly opposite Anthropic's own security tooling. https://techcrunch.com/2026/06/22/openai-launches-new-initiative-to-help-find-and-patch-open-source-bugs/ - Groq confirms a $650M raise (led by Disruptive and Infinitum) and a rebuilt executive bench after Nvidia's ~$20B "not-acqui-hire" in December licensed Groq's tech and hired away founder Jonathan Ross and president Sunny Madra. Groq pivots to its neocloud — 13 data centers, 5M+ developers, trillions of tokens a week. https://techcrunch.com/2026/06/22/ai-chipmaker-groq-confirms-650m-raise-re-staffs-after-nvidias-20b-not-acqui-hire-deal/ - SpaceX signs a compute deal worth up to $6.3B with open-weight lab Reflection AI — $150M/month from July 1, 2026 through 2029 at the former xAI "Colossus" data center near Memphis. Context: Anthropic pays SpaceX ~$1.25B/month and Google ~$920M/month. https://techcrunch.com/2026/06/22/spacex-inks-compute-deal-with-reflection-ai-an-open-source-ai-lab/ Quick hits - Google DeepMind invests $75M in indie studio A24 to build AI filmmaking tools; Demis Hassabis frames it as supporting "authentic, meaningful storytelling." https://techcrunch.com/2026/06/22/google-deepmind-bets-75m-on-ais-future-in-hollywood-with-a24-deal/ - Anthropic brings the full Claude Desktop app (chat, Cowork, and Claude Code) to AWS, Google Cloud, and Microsoft Foundry, with in-cloud inference and enterprise SSO/policy controls. Hanwha Solutions deployed it to hundreds of users. https://claude.com/blog/the-full-claude-desktop-experience-on-aws-google-cloud-and-microsoft-foundry - Oracle cut ~21,000 jobs over twelve months (13% of staff) and explicitly cited AI adoption in its filing, joining a 2026 list that includes Meta, Intuit, Cisco, GitLab, and Cloudflare. https://techcrunch.com/2026/06/22/the-running-list-major-tech-layoffs-in-2026-where-employers-cited-ai/ - Salesforce agrees to acquire AI customer-service platform Fin for $3.6B; Fin's "Apex" model claims to resolve ~76% of support volume end-to-end, folding into Agentforce. https://techcrunch.com/2026/06/15/salesforce-acquires-ai-customer-service-platform-fin-for-3-6b/
    続きを読む 一部表示
    7 分