Friday, August 7, 2026
Sign in to LinkedIn, then select Subscribe.
Foundational AI Updates
Jeff Dean and three senior Google researchers leave to launch Discovery Loop
On August 5, 2026, TechCrunch reported that Jeff Dean — Google's chief scientist, who joined in 1999 as roughly the company's 30th employee — is leaving alongside senior fellow Sanjay Ghemawat, Google Brain founding member Quoc Le, and Google DeepMind senior research scientist Oriol Vinyals to found Discovery Loop, a public benefit corporation. The venture's stated aim is to use AI to accelerate scientific research by automating experimental loops and enabling recursive self-improvement of the underlying models. Radical Ventures and Khosla Ventures are co-leading the financing, with Kleiner Perkins, Lightspeed and Doerr Capital participating, and Alphabet also lending support. Dean framed the thesis in terms of throughput rather than raw model capability: "You will get both a higher quantity and a higher quality of experiments, and that will lead to scientific breakthroughs and advances."
Why it matters: A heavily capitalized lab aimed squarely at automating the experiment-to-next-experiment loop is targeting the same bottleneck that governs preclinical and translational science, so R&D leaders should expect a new class of vendor pitching closed-loop discovery rather than another target-prediction model. The Alphabet-adjacent pedigree also makes it likely the tooling arrives through cloud relationships large sponsors already hold.
Source: TechCrunch — Jeff Dean and other top AI researchers are leaving Google to launch their own startup
Anthropic begins building an in-house custom AI chip design team
On August 5, 2026, TechCrunch reported that Anthropic has started recruiting engineers with chip design experience for a custom silicon team, with the stated goal of co-designing hardware and models so its technology runs faster and more efficiently. The move follows the same path as its largest rivals: OpenAI unveiled its Broadcom-built Jalapeño inference chip in June 2026, Google DeepMind runs on Alphabet's TPUs, and Meta is developing its MTIA accelerators for AI workloads. Anthropic currently reaches compute through partnerships with AWS, Google, Nvidia and AMD rather than through silicon of its own. The Information has separately reported that the company has explored Samsung as a potential manufacturing partner for the effort.
Why it matters: When frontier labs move to bespoke silicon, it signals that inference cost rather than model quality is becoming the binding constraint on large-scale deployment — a caution for sponsors whose multi-year AI budgets assume per-token prices keep falling. It also deepens supplier lock-in, because models tuned to proprietary hardware are harder to port if a vendor relationship changes.
Source: TechCrunch — Anthropic is hiring an AI chip design team
Drug Development & AI
Genentech options its first neuroscience target from Recursion's AI-mapped genome screen
In second-quarter results released on August 5, 2026, Recursion Pharmaceuticals said Genentech had exercised the first Validated Target Option under their neuroscience collaboration, moving a previously unexplored target into small-molecule early discovery. The target emerged from a whole-genome CRISPR knockout map built with Roche and Genentech from more than one trillion iPSC-derived neuronal cells, with machine-learning predictions then carried through jointly designed pathway, functional and disease validation stages. Recursion has now received $216 million in upfront and milestone payments under the partnership, which carries potential development and commercialization milestones exceeding $300 million per program. The company reported $556.8 million in cash as of June 30, a net loss of $131.0 million for the quarter, and cut 2026 cash operating expense guidance to under $375 million, extending runway into early 2028. It also said newly deployed agentic AI tools reduced one structural analysis step from four hours to 30 minutes and contributed to enrollment rates 1.3 to 1.6 times historical benchmarks.
Why it matters: This is one of the first cases of a large pharma partner formally validating and optioning a target that a computational platform surfaced rather than a human hypothesis-led program — the step that separates platform revenue from platform proof. Sponsors evaluating AI discovery partners should be asking for option-exercise counts and validation-stage attrition, not model benchmarks.
Lunai and BrainStorm pair AI patient stratification with Parkinson's brain organoids
On August 4, 2026, Lunai Bioworks said its BioSymetrics subsidiary had signed a master collaboration agreement with BrainStorm Therapeutics to apply AI-based patient stratification to Parkinson's disease, with an initial data readout targeted for the first half of 2027. BioSymetrics' Contingent AI platform will analyze clinical datasets to resolve Parkinson's into biologically distinct subtypes, and BrainStorm will test the resulting hypotheses in patient-derived familial Parkinson's midbrain organoids paired with AI foundation models. An earlier BioSymetrics white paper drawing on the Parkinson's Progression Markers Initiative dataset had already separated subgroups associated with fast motor progression, cognitive decline, and a female-enriched profile. Lunai chief executive David Weinstein said Parkinson's "is not one biological entity and treating it as a single disease has contributed to repeated clinical-development challenges," while BrainStorm's Robert T. Fremeau Jr. described the work as "an iterative path from patient stratification to human biological validation."
Why it matters: Enrolling a biologically heterogeneous population against a single endpoint is a recurring explanation for neurology trial failures, and this design tests a computational subtype hypothesis in human tissue before a protocol is written. CNS sponsors should watch whether the organoid step actually discriminates the subtypes, because that is precisely what regulators will probe if stratification is later used to define eligibility.
Regulatory, Policy & Industry Moves
FDA and CMS ran an unannounced clinical AI demo day with ten companies
On August 5, 2026, STAT reported that the FDA and the Centers for Medicare and Medicaid Services have been holding closed-door meetings with industry on clinical AI, including a "clinical AI demo day" at the FDA's White Oak headquarters where officials tried AI clinical tools firsthand and discussed how such products should be regulated and paid for under Medicare. Ten companies took part: Anthropic, Counsel Health, Curai, K Health, Microsoft AI, Amazon One Medical, Doctronic, Ellipsis Health, Hippocratic AI and Welldoc. STAT interviewed four attendees — Counsel Health, K Health, Hippocratic AI and Ellipsis Health — and characterized the sessions as conversations that had not been publicly announced. The report noted the format handed a mix of large technology firms and venture-backed startups, including companies funded by Andreessen Horowitz and Khosla Ventures, an early hand in shaping policy.
Why it matters: Coverage and evidence expectations for clinical AI are being scoped in rooms that drug sponsors and their evidence-generation teams are not in, and those expectations will eventually govern AI-enabled endpoints and decision-support claims. Teams planning such submissions should treat the public docket as a lagging indicator and seek direct agency engagement rather than waiting for guidance.
UK regulator draws the medical-device line for AI scribes used in the NHS
On August 4, 2026, STAT covered the UK Medicines and Healthcare products Regulatory Agency's decision setting out when ambient voice technology — AI scribes — is regulated as a medical device in Great Britain. Under guidance issued with NHS England, AVT products used solely to transcribe consultations, summarize clinical conversations, draft correspondence or suggest clinical codes for a clinician to review are not medical devices. Products that support diagnosis, treatment or disease prevention decisions, or that can take automated action without clinician oversight, remain regulated as devices. MHRA chief executive Lawrence Tallon said the guidance is "an important step forward for the safe and confident adoption of AI in healthcare settings," describing the intent as removing ambiguity about which functions cross the regulatory line.
Why it matters: This is the sharpest regulatory articulation yet of the boundary between documentation-support AI and clinical-decision AI, and it is the same boundary sponsors face when generative tools draft source documents, safety narratives or medical-review summaries. Vendors now have a commercial incentive to keep features on the unregulated side of that line, so sponsors should qualify what a tool actually does rather than how it is positioned.
Source: STAT — Are AI scribes medical devices? U.K. regulator weighs in
Originally published on LinkedIn. Read the original
