[{"data":1,"prerenderedAt":12},["ShallowReactive",2],{"$f2owquy72czfiu":3},{"slug":4,"title":5,"date":6,"dateDisplay":7,"minuteRead":8,"description":9,"problems":10,"html":11},"2026-08-02_2026-08-08","Jeff Dean and three colleagues leave Google","2026-08-09","August 9, 2026",5,"DeepMind changes hands; four senior researchers leave. Google announced on August 5 that Demis Hassabis moves from running Google DeepMind to chairing it…",[],"\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">01\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">DeepMind changes hands; four senior researchers leave\u003C\u002Fh2>\n  \u003Cp>Google announced on August 5 that Demis Hassabis moves from running Google DeepMind to chairing it, handing the lab to its CTO, Koray Kavukcuoglu. Hassabis adds Alphabet chief scientist to his titles and keeps running Isomorphic Labs, the drug-discovery company. Four senior Google researchers announced the same day that they were leaving to found Discovery Loop, a public benefit corporation: Jeff Dean, at the company since 1999; Sanjay Ghemawat, his longtime collaborator; Oriol Vinyals, a DeepMind research VP; and Quoc Le, a cofounder of Google Brain. The company says it will automate whole experimental loops — running and iterating thousands of experiments — in areas like drug discovery and chip design. Alphabet is a founding investor and cloud partner. \u003Ca href=\"https:\u002F\u002Fx.com\u002FJeffDean\u002Fstatus\u002F2085034604172603724\" rel=\"noopener\">Dean\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fblog.google\u002Fcompany-news\u002Finside-google\u002Fmessage-ceo\u002Fnext-chapter-ai-momentum\u002F\" rel=\"noopener\">Google\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">02\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Humans approving agent commands miss the dangerous ones\u003C\u002Fh2>\n  \u003Cp>A browser game imitating Claude Code's permission prompts logged 409,000 decisions and found players approving one in three malicious commands — under a 60-second timer and a threat rate far above real life, by the builder's own account. Anthropic's controlled experiment found worse: 1,053 paid testers caught 13.6% of planted dangerous commands. On August 14, Claude Code defaults new Pro, Max, and Team sessions to a classifier gate that catches 89% by Anthropic's own evaluation — not a substitute for a sandbox. \u003Ca href=\"https:\u002F\u002Fscalex.dev\u002Fblog\u002Fai-agent-permissions-stats\u002F\" rel=\"noopener\">Game data\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fclaude.com\u002Fblog\u002Fauto-mode-default-in-claude-code\" rel=\"noopener\">Anthropic\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">03\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Sixteen of roughly 300 AI-designed viruses were viable\u003C\u002Fh2>\n  \u003Cp>Arc Institute and Stanford researchers prompted two genome models with a bacteriophage — a virus that infects bacteria, with one of biology's simplest genomes — and got roughly 300 candidate genomes, 16 of which assembled into viruses that kill laboratory E. coli. That is the range shown: one template, lab strains, dishes. The models were kept off human viruses by leaving those sequences out of the training data, an exclusion the authors say is reversible; the weights are public. Preprinted last September, in Science on August 6. \u003Ca href=\"https:\u002F\u002Fwww.science.org\u002Fdoi\u002F10.1126\u002Fscience.aec2657\" rel=\"noopener\">Science\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--paragraph\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">04\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Memory demand from AI outruns supply through 2027\u003C\u002Fh2>\n  \u003Cp>Memory makers are shifting wafer capacity to high-bandwidth memory — the stacked DRAM that feeds AI accelerators — and the rest of the market is repricing around them. \u003Ca href=\"https:\u002F\u002Fwww.trendforce.com\u002Fpresscenter\u002Fnews\u002F20260105-12860.html\" rel=\"noopener\">TrendForce\u003C\u002Fa>, the memory-market tracker, revised its Q1 2026 estimate for conventional DRAM contract prices to a 90–95% rise. The surge is now impacting retail: Microsoft \u003Ca href=\"https:\u002F\u002Fnews.xbox.com\u002Fen-us\u002F2026\u002F06\u002F25\u002Fxbox-console-price-update\u002F\" rel=\"noopener\">raised Xbox prices\u003C\u002Fa> $150 on the top model on August 1 citing memory and storage costs, Sony raised PS5 prices in April, and Valve says its Steam Machine launched about $300 over its planned price — the same stated cause each time. Relief is not close: Micron told investors its HBM is \u003Ca href=\"https:\u002F\u002Fwww.investing.com\u002Fnews\u002Ftranscripts\u002Fearnings-call-transcript-micron-tops-q3-2026-estimates-shares-jump-146-93CH-4759504\" rel=\"noopener\">booked through 2027\u003C\u002Fa> behind $22 billion in non-cancelable agreements, Samsung expects the 2027 gap to widen, and no meaningful new capacity lands before 2028. The supply side's escape valve: SK hynix and Sandisk published the first \u003Ca href=\"https:\u002F\u002Fnews.skhynix.com\u002Fen\u002Fhbf-at-fms-2026\u002F\" rel=\"noopener\">High Bandwidth Flash spec\u003C\u002Fa> on August 4 — a flash tier below HBM for inference, up to 3 TB\u002Fs on paper, with no product behind it yet.\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--blurbs\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">05\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">Companies began budgeting their own AI token spend\u003C\u002Fh2>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">Microsoft told engineers to stop maximizing token use.\u003C\u002Fstrong> An internal August 4 memo from Jay Parikh, who runs the CoreAI division, says \"tokenmaxxing is not what we are optimizing for\" and makes a cheaper OpenAI model the default in GitHub Copilot for staff. Divisions have run under token budget targets since July; CoreAI has set no caps for individual teams or engineers. \u003Ca href=\"https:\u002F\u002Fwww.404media.co\u002Fmicrosoft-tells-engineers-tokenmaxxing-is-not-what-we-are-optimizing-for\u002F\" rel=\"noopener\">404 Media\u003C\u002Fa>\u003C\u002Fp>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">SAP paused most hiring and non-AI travel to afford its AI push.\u003C\u002Fstrong> An \u003Ca href=\"https:\u002F\u002Fwww.404media.co\u002Fsoftware-giant-sap-stops-most-travel-and-hiring-because-of-ais-soaring-cost\u002F\" rel=\"noopener\">executive-board email\u003C\u002Fa> in early July limited new hires at Europe's largest software company to mainly core AI roles and stopped internal travel unrelated to AI. SAP calls that moving money toward AI rather than AI costs running away; \u003Ca href=\"https:\u002F\u002Fwww.investing.com\u002Fnews\u002Fcompany-news\u002Fsap-q2-2026-slides-cloud-backlog-surges-27-amid-margin-pressure-93CH-4810220\" rel=\"noopener\">coverage of its Q2 presentation\u003C\u002Fa> names token costs among the drags on margin.\u003C\u002Fp>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">At Rippling, a tenth of employees ran most of the bill.\u003C\u002Fstrong> At the payroll software company, 10-15% of staff drove about 60% of AI spend. One engineer ran $50,000 a month. Rippling has since announced the metering dashboard it built in response, now open for a waitlist. \u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F07\u002Fafter-rippling-blew-millions-on-ai-in-months-it-built-an-employee-roi-tool\u002F\" rel=\"noopener\">TechCrunch\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--blurbs\">\n  \u003Cspan class=\"section__marker\" aria-hidden=\"true\">06\u003C\u002Fspan>\n  \u003Ch2 class=\"section__heading\">What AI crawling returns depends on the site\u003C\u002Fh2>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">A content site served 214 bot page loads for every human pageview.\u003C\u002Fstrong> A 1.5-million-page database of charitable donors served 1.28 million pages to automated clients in the week to August 7, against 5,977 pageviews from people. Anthropic's search crawler alone took 420,680 pages and sent twelve visitors back. Both counts come from the operator's own dashboards. \u003Ca href=\"https:\u002F\u002Fpatronview.com\u002Fnews\u002F99-percent-of-my-website-traffic-is-bots\u002F\" rel=\"noopener\">The logs\u003C\u002Fa>\u003C\u002Fp>\n  \u003Cp class=\"blurb\">\u003Cstrong class=\"blurb__lead\">Retail sites report the traffic running the other way.\u003C\u002Fstrong> Shopify told investors on August 5 that AI-driven traffic and orders to its merchants each tripled year over year — multiples on a base it declined to disclose. Adobe, across more than a trillion US retail visits, measures AI-referred traffic converting 54% better, a relative gain; a year earlier it converted worse. \u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F08\u002F05\u002Fshopify-says-ai-search-is-driving-more-traffic-and-sales-not-replacing-google\u002F\" rel=\"noopener\">TechCrunch\u003C\u002Fa>, \u003Ca href=\"https:\u002F\u002Fwww.digitalcommerce360.com\u002F2026\u002F08\u002F04\u002Fadobe-ai-referred-traffic-retail\u002F\" rel=\"noopener\">Adobe\u003C\u002Fa>\u003C\u002Fp>\n\u003C\u002Fsection>\n\u003Csection class=\"section section--further-reading\">\n  \u003Cimg class=\"section__mark\" src=\"\u002Fbranding\u002Fstargazer.png\" width=\"54\" height=\"25\" alt=\"\" \u002F>\n  \u003Ch2 class=\"section__heading\">Something to chew on\u003C\u002Fh2>\n  \u003Cp class=\"read\">\u003Ca class=\"read__title\" href=\"https:\u002F\u002Fwww.noemamag.com\u002Fwhy-is-everyone-in-tech-so-sad\u002F\" rel=\"noopener\">Why Is Everyone In Tech So Sad?\u003C\u002Fa> — Aaron Horwath, who implements AI at a creative technology company for a living, argues that the cost nobody budgets for is what an organization becomes once the people who stayed for the colleagues and the craft decide the work no longer supplies either.\u003C\u002Fp>\n  \u003Cp class=\"read\">\u003Ca class=\"read__title\" href=\"https:\u002F\u002Fwww.platformer.news\u002Freplacing-casey-with-claude\u002F\" rel=\"noopener\">How much of my boss's job can AI do?\u003C\u002Fa> — A reporter at the tech newsletter Platformer built a model of her editor from six years of the archive, his edits on her drafts, and a year of team Discord: it wrote passably and edited badly, 70% of its notes off the mark by her count against 5% for him.\u003C\u002Fp>\n\u003C\u002Fsection>",1786309605550]