In partnership with
In Today’s Issue:
🧠 Hassabis steps back and four architects leave Google
🔧 Anthropic starts designing its own chips
🕵️ Meta's test agent breaks out and hacks a company
📊 Muse Spark 1.2 crashes the coding top group
✨ And more AI goodness…
⚡ The Signal
Google reorganized the top of its AI operation this week and came out of it with more titles and fewer builders.
Demis Hassabis becomes chair of Google DeepMind and chief scientist of Alphabet, which sounds like elevation and also means distance from the shipping line. On the same day, Jeff Dean and three colleagues who between them built MapReduce, TensorFlow, the TPU line and much of Gemini walked out to start a company whose stated purpose is to automate scientific discovery. The rest of today's issue rhymes with that. Anthropic is hiring chip engineers because renting compute is no longer enough, Meta is training its ad model on frontier-scale infrastructure and shipping its first coding agent, and the UK's security institute spent the week cataloging what those agents do when they get loose on the open internet. The field is finishing its move from research project to industrial operation, and the researchers have noticed.
All the best,
Kim Isenberg
GEM, Meta's hybrid recommendation and LLM architecture for ads.
💰 Meta Trains Its Ad Brain Like a Frontier Model
Meta has published how it trains GEM, the foundation model behind ad recommendations on Facebook and Instagram, and the engineering reads like frontier LLM work.
GEM runs on several thousand of Meta's latest-generation GPUs and carries trillions of sparse parameters alongside billions of dense ones. Meta says it doubled end-to-end training efficiency to 20 to 25 percent MFU (model FLOPs utilization, the share of a GPU's theoretical peak that a training run actually uses) and scaled training compute fourfold in twelve months.
👉 tl;dr: The scaling playbook that builds chatbots is now pointed at the ad system that pays for them.
Anthropic is building an in-house silicon team for Claude.
🔧 Anthropic Starts Designing Its Own Chips
Anthropic has confirmed for the first time that it is building an in-house silicon team to design custom chips for Claude.
The lab is hiring chip engineers on a listed range of $320,000 to $485,000, and a spokesperson said Anthropic will co-design hardware and models so Claude runs faster and more efficiently "at the scale our customers need." Anthropic stresses that this stays a "multi-chip approach", with AWS, Google, Nvidia and AMD still central to its scaling. The Information reported in July that the lab has held talks with Samsung as a possible manufacturing partner.
👉 tl;dr: Custom silicon is becoming table stakes at the frontier, and Anthropic is the latest to move, after OpenAI's Broadcom-built Jalapeno and Meta's next-generation chip due in September.
(Jeff Chiu / AP via The Information)
🕵️ Meta's Test Agent Broke Out and Hacked a Company
Meta's Muse Spark 1.1 escaped its testing sandbox and altered another company's internal systems, which makes Meta the third frontier lab in recent weeks to lose containment of an evaluation agent.
Meta and its testing partner Irregular blame a misconfigured sandbox, which Irregular calls "the exact same evaluation-environment issue that was already disclosed by Anthropic last week." The harder news landed on Tuesday, when the UK AI Security Institute reported that across 122 test runs, agents took unauthorized action against real people and organizations in ten of them, and one agent invented fake online identities to pressure an open-source maintainer into approving malicious code.
👉 tl;dr: Sandbox escapes are becoming routine. An agent building sock puppets to manipulate a human is new.
🎬 Watch This
Reed Hastings cofounded Netflix, sits on Anthropic's board, and has spent years funding education, which makes him an unusually well-placed witness to what AI does to both. At TED2026 he talks with Sal Khan about the race to build AI, why he thinks a personal tutor for every student is now within reach, and why the economic disruption arrives faster than most people expect. (Recorded April 15, 2026)
"good iteration for muse spark 1.2 onto bigger and better things "
– Alexandr Wang, Chief AI Officer at Meta and head of Meta Superintelligence Labs
Wang posted this on Wednesday, hours after shipping Muse Code, the terminal agent that runs on Muse Spark 1.2 through the Meta Model API. That watermelon is a tease: Watermelon is Meta's internal codename for its next flagship model, still in training, which Wang told staff in July had already caught GPT-5.5. Today's Graph shows why he is looking past 1.2 already.
A federal appeals court just handed Perplexity its Amazon fight back. The Ninth Circuit vacated the injunction on Comet's shopping agent, reasoning that under the federal anti-hacking law it is the user, not Perplexity, who accesses Amazon's servers.
The Perplexity app icon on a phone screen. (Engadget)
Hassabis Steps Back, and Jeff Dean Takes Three Architects With Him
The Takeaway
👉 Demis Hassabis hands over day-to-day control of Google DeepMind to become its chair and chief scientist of Alphabet, keeping Isomorphic Labs.
👉 Koray Kavukcuoglu rises from CTO to SVP, taking over Gemini model development, frontier research, and the Gemini app and developer teams.
👉 Jeff Dean leaves after 27 years and takes Sanjay Ghemawat, Quoc Le and Oriol Vinyals with him to Discovery Loop, a new public benefit corporation.
👉 Google is a founding investor and cloud partner in the startup, but Alphabet still slid about 5 percent on the news.
Google spent Wednesday rearranging the top of its AI organization, and when it was done the company had kept its most famous scientist and lost four of the engineers who built Google. Demis Hassabis is handing over day-to-day control of Google DeepMind to become its chair and chief scientist of Alphabet. The same announcement confirmed that Jeff Dean is leaving after 27 years, and that he is not leaving alone.
(Demis Hassabis presenting "a new golden age of scientific discovery." Google via the-decoder)
Hassabis presents his own move as stepping up rather than stepping away. "I've decided that now is the right time for me to hand over my day-to-day operational responsibilities at GDM," he wrote, describing a shift toward the long arc of AGI and a strategic remit alongside Sundar Pichai. He keeps Isomorphic Labs, Alphabet's drug discovery spinout. Operational control of Gemini passes to Koray Kavukcuoglu, promoted from chief technology officer to senior vice president, who now owns model development, frontier research, and the Gemini app and developer teams.
The second half of the announcement is the heavier one. Jeff Dean, at Google since 1999, is co-founding Discovery Loop with Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Between them the four are responsible for MapReduce, BigTable, Spanner, TensorFlow, the TPU line, Google Brain, sequence-to-sequence learning, AlphaStar, and the Gemini models Vinyals co-led until this week. Their own founding deck notes they have "worked together for 10 to 30 years." The company is a public benefit corporation, and its stated purpose is to "automatically solve important problems in machine learning, science, and engineering."
(Discovery Loop's co-founders: Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Discovery Loop)
Google is not being frozen out. It is a founding investor and Discovery Loop's cloud partner, which keeps the split amicable and the compute bill on Google's own ledger. The direction of travel is still unmistakable. Four people who spent a quarter of a century building the substrate of modern AI inside a very large company have concluded that the next phase is better built outside one, which is the conclusion Yann LeCun reached about Meta in late 2025. Alphabet slid on the news, at one point down about 5 percent and briefly some $190 billion lighter, before recovering.
The skeptical reading deserves room too. A chair-and-chief-scientist title can mean elevation or it can mean distance from the shipping line, and Google has not said which. A research quartet with a mission statement is also not yet a product. Discovery Loop has four famous founders, an investor and a name, but no model.
Why it matters: Google's standing answer to the AI race has always been that it employs the people who invented the field. This week it employs fewer of them, and the ones who left are betting that automating scientific discovery is a job for a startup rather than for the company that has been promising a golden age of it from its own keynote stage.
Sources:
🔗 https://blog.google/company-news/inside-google/message-ceo/next-chapter-ai-momentum/
🔗 https://t.co/Rv3LMdLluK
🔗 https://the-decoder.com/google-deepmind-loses-both-its-ceo-and-chief-scientist-as-demis-hassabis-and-jeff-dean-step-down-simultaneously/
The GTM Playbook Behind Warmly's Acquisition
Warmly ran pipeline, outreach, and lead scoring on autopilot for hundreds of startups — before a single sales hire.
HubSpot acquired them for it. Now the cofounders are walking you through the exact system, live, before they disappear into product. Join the Builder Session on August 12.
Leave with an agentic GTM stack you can replicate this week. Plus HubSpot Credits when you join HubSpot for Startups.
The chart: Meta's own numbers for Muse Spark 1.2, the coding model it shipped on Wednesday, across three agent benchmarks. On Terminal-Bench 2.1 (89 real terminal tasks), Opus 5 at max effort leads with 86.7%, ahead of Muse Spark 1.2 at 82.9% and GPT-5.6 Terra at 81.8%. On DeepSWE 1.1, Opus 5 takes 65.0% and Muse Spark 1.2 slips to third at 59.3%. On Meta's internal coding bench, Opus 5 leads again at 79.4%, with Muse Spark 1.2 second at 70.6%.
The lesson: Meta's first serious run at coding agents lands it in the top group on day one. It tops nothing on this chart, but the distance to Anthropic's leader is now single digits rather than tiers.
The caveat: These are Meta's own runs, and each model is scored inside a different harness (Claude Code, Codex, Grok Build, Muse Code), so the bars measure model plus scaffold, not model alone. The OpenAI comparison also runs against GPT-5.6 Terra rather than the stronger Sol tier.
🧠 What Ozempic Might Be Doing to the Brain
⚡ Bottom line: A review of 30 preclinical studies found GLP-1 drugs cut amyloid and tau across four different agonists.
💡 Why it matters: The weight-loss drugs millions already take may act on Alzheimer's core pathology, not only on body weight.
🔎 What it means: The animal evidence is strong and the human evidence is thin, and that gap is the whole story.
Alzheimer's research has spent decades trying to clear two kinds of debris from the brain: amyloid-beta, which clumps into plaques between neurons, and tau, which tangles inside them. A systematic review in Molecular and Cellular Neuroscience, led by Simon Cork of Anglia Ruskin University, asked whether a drug class already sitting in millions of fridges is quietly doing that job.
Ozempic semaglutide injection pens. (PsyPost)
Across 30 preclinical studies covering four drugs, the pattern is unusually consistent. Liraglutide has the strongest record, reducing amyloid-beta in 13 of 15 experiments and tau in all 12 that measured it. Exenatide cut amyloid in 6 of 8, semaglutide in 3 of 4, and dulaglutide reduced accumulation in both studies that tested it while improving learning and memory in mice.
The review proposes four routes, and only the first is about clearing the mess. GLP-1 drugs appear to suppress BACE1, the enzyme that cuts amyloid-beta out of a larger protein, so less plaque material gets made in the first place. They also improve insulin sensitivity in the brain, which is what keeps tau from tangling, reduce inflammation, and improve cardiovascular health. Think of it as fixing the factory rather than sweeping the floor.
The four mechanisms by which GLP-1 receptor agonists may act on Alzheimer's pathology (Cork et al., Molecular and Cellular Neuroscience, via MedicalXpress)
Then comes the gap that matters. The human evidence is two small trials, 38 patients on liraglutide and 21 on exenatide, and neither delivered meaningful cognitive benefit. Lab mice do not reproduce the widespread neuron death or the decades of aging that define the human disease. Cork's own framing stays careful: "with more than three-quarters of preclinical studies showing reductions in amyloid-beta or tau, and early signals emerging from studies on humans, GLP-1 drugs remain strong candidates for future Alzheimer's prevention trials." Prevention rather than reversal, and trials rather than prescriptions.
The toughest room in advertising. Ad Studio just cleared it.
Major brands just took over a Times Square billboard, and every ad was built in Hightouch Ad Studio. Sweetgreen, HelloFresh, and Tripadvisor cleared the toughest room in advertising. Create, edit, and launch on-brand ads in a few clicks.