Wednesday, September 16, 2026

Pulse (last 3d) · Research 94 · Agents 48 · Models 33 · Products 26 · Industry 22 · Policy 20 · Open Source 15 · Regulation 15 · Legal 13 · Infra 10

Trending · OpenAI 14 · Anthropic 10 · Dario Amodei 5 · Donald Trump 5 · FDA 5 · Nvidia 5 · Claude 4 · Hugging Face 4 · Jensen Huang 4 · Apple 3 · Claude Code 3 · GitHub 3

---

• Nvidia is acquiring Hugging Face for $12.93B, putting the default open-weights distribution layer inside the dominant hardware vendor. blogs.nvidia.com
• Merck and Moderna's intismeran autogene plus Keytruda hit both RFS and DMFS in resected stage IIB-IV melanoma, the first positive Phase 3 for an individualized neoantigen therapy. merck.com
• Generate Biomedicines registered GB-4362, an AI-designed antibody whose job is to mop up free MMAE payload so other people's ADCs can be dosed higher. clinicaltrials.gov
• Sakana's PC-ALM trains 1000-layer networks with layer-local learning and no backpropagation, by turning each layer into a PI feedback controller. pub.sakana.ai
• FDA is running two parallel authorization tracks for generative-AI medical devices, UpDoc via 510(k) and RecovryAI via De Novo, which means the pathway you get depends on your predicate story. hlth.com
• Coding agents have converged hard enough that SWE-bench can no longer order its top entries, per a new arXiv preprint proposing replacements. arxiv.org
• Repository-scale vulnerability localization is still unsolved: GPT-5.5 at maximum reasoning effort scores 0.221 F1 on a new benchmark. x.com

Four separate readouts stop making sense as individual drug stories and start making sense as one story about designed biology being used to fix the constraints of adjacent modalities.

Generate's GB-4362 is a designed antibody whose product is another drug's therapeutic window Generate Biomedicines registered NCT07484022 for GB-4362, an AI-designed anti-MMAE antibody built to scavenge free payload shed from antibody-drug conjugates in circulation. The framing in their own write-up is explicit: design past the tolerability limit rather than design a better warhead. That inverts the usual generative-biology pitch, where the model produces the asset; here the model produces the enabling layer that makes a whole class of existing assets dosable higher. If you are building models over oncology data, this is the shape to expect more of: the valuable prediction is not "will this molecule bind" but "what constraint in the existing regimen can be engineered away," which is a systems question and needs combination-level data your pipeline probably is not currently joined on. clinicaltrials.gov | generatebiomedicines.com | biospace.com

Individualized neoantigen therapy clears Phase 3 for the first time Merck and Moderna report that intismeran autogene plus pembrolizumab met both recurrence-free and distant-metastasis-free survival endpoints against pembrolizumab alone in resected stage IIB-IV melanoma. This is a company press release with topline only, no published curves and no independent replication, so treat effect sizes as unstated. What matters regardless of magnitude is that the manufacturing-and-sequencing loop per patient survived a registrational trial, which is the part everyone assumed would break. For anyone running a precision-medicine data platform, this converts per-patient variant-to-manufacturing turnaround from a research workflow into a regulated one, with the provenance, versioning and audit obligations that implies. merck.com

The ADC field now has overall-survival benefit and a crowded lane at the same time A B7-H3 ADC has posted the first Phase 3 overall-survival benefit in that target class, while Kelun-Biotech's sacituzumab tirumotecan plus pembrolizumab met its primary endpoint in first-line PD-L1-negative non-squamous NSCLC, and BioNTech is bringing first global data for a PD-(L)1 x VEGF bispecific combined with an ADC to WCLC 2026 as a late-breaking oral. All three are company-issued releases ahead of full presentation. Taken together, ADCs have moved from "does the modality work" to "which target, which payload, which combination partner," which is a selection problem rather than a discovery one. That is a direct argument for investing in target-and-combination ranking models over molecular-property models, because the bottleneck has moved. investis.com | prnewswire.com | biontech.com

In vivo delivery is quietly becoming the platform layer under both CAR-T and gene editing CRISPR Therapeutics reported Phase 1a data for CTX310 showing deep and durable ANGPTL3 editing with triglyceride and LDL lowering at ESC 2026, and its business update frames in vivo CAR-T for autoimmune disease as a diversifying platform rather than a single program. The common factor across both is that the engineering effort has shifted from the editor or the receptor to the delivery vehicle and the tissue targeting. For data teams, that means the covariates that predict success are increasingly pharmacokinetic and tissue-distribution variables, not sequence features, and most oncology-trained feature stores do not carry them. globenewswire.com | crisprtx.com

Three independent items today add caveats, propose replacements, or document their own training run, and the reason is the same: the headline number stopped discriminating.

Insilico's O3DC makes "known caveats" a required field in a benchmark index Insilico has convened O3DC, an open index of drug-discovery benchmarks whose novel design choice is that each entry carries its own caveats alongside its scores. That is a governance move disguised as a data structure: it makes the failure mode of a benchmark citable, which is the only thing that stops a number travelling further than its validity. Whether it gets adopted depends entirely on whether the caveat field stays populated after the launch cohort, and there is no evidence yet either way. If you evaluate external AI-drug-discovery vendors, this is the first artifact you can point at when asking a vendor to state the caveats of the benchmark they just quoted at you. o3dc.org

SWE-bench can no longer rank its own leaders, and a preprint says what to measure instead A new arXiv submission argues coding agents have converged tightly enough at the top of SWE-bench that the leaderboard cannot order its top entries, and proposes alternative measurements. This is a preprint with no independent replication, but it matches what anyone running agents in production already sees: the reported delta between the top models does not predict which one finishes your actual ticket. Pair it with the vulnerability-localization result below and the picture is that saturated benchmarks are hiding very large remaining gaps in the tasks that matter. If you are choosing a coding model for an agent squad, stop using leaderboard rank as the tiebreaker and build a small internal eval on your own repository shape. arxiv.org

Repository-scale vulnerability finding is nowhere near solved: 0.221 F1 at maximum effort A newly released Vulnerability Localization Benchmark reports that GPT-5.5 at its highest reasoning effort achieves 0.221 F1 at finding vulnerable code across a whole repository. The claim comes via an author tweet rather than a reviewed paper, so treat the exact figure as provisional, but the order of magnitude is the point. Repository-scale security review is a task everybody's agent roadmap assumes is nearly free, and it is not. If your agent squads are heading toward autonomous code changes on clinical-data services, this number is the argument for keeping a human security gate in the loop and for not letting an agent's own scan stand as the control. x.com

OPEN-1B publishes a fully auditable end-to-end training run OPEN-1B documents a complete 1B-parameter training run with the entire pipeline auditable, which is the transparency counterpart to the benchmark-caveat move. The value is not the model, which is small, but the existence of a reference artifact showing what full provenance for a training run actually costs to produce. For a regulated environment where you will eventually be asked to evidence how a model was trained, this is worth reading as a template rather than as a release. arxiv.org

Four results from very different places all move inference off the assumed datacenter default, and together they change the cost model for deploying anything near patient data.

PC-ALM trains 1000 layers with no backpropagation at all Sakana AI's PC-ALM adds per-layer Lagrange multiplier "dual neurons" to predictive coding, turning each layer into a PI feedback controller, and reports the first layer-local learning method demonstrated to depths of 1000 layers. Layer-local means no global backward pass, so the memory and synchronization assumptions that force training onto tightly-coupled accelerator clusters weaken. This is a lab publication on its own site without independent replication, and depth is not the same as quality, so hold the result loosely. But if layer-local training holds up, training on federated or physically separated clinical data stops being a privacy workaround and becomes an architecture, which is a materially different conversation with a hospital partner. pub.sakana.ai

A 44M-parameter ternary model ships in 19.8 MB and runs at ~1,900 tok/s on a laptop CPU A hobbyist trained a 44M-parameter quantized LLM from scratch on 45B tokens; it ships as a single 19.8 MB file, runs around 1,900 tokens per second on a laptop CPU and about 500 tok/s in a browser tab via WASM. Self-reported, unreplicated, and a 44M model does not do most of what you want. The number that matters is not the quality, it is 19.8 MB, which is small enough to embed inside an application rather than call. Meanwhile JustFit demonstrates 200K-token serving on a 24 GiB laptop via just-in-time state management, attacking the same problem from the context side. For anything where the data cannot leave a site, the question is shifting from "can we get a model in there" to "which tier of model do we put in there." reddit.com | arxiv.org

Gemini 3.8 Live replaces the ASR-LLM-TTS pipeline with native speech-to-speech Google shipped Gemini 3.8 Live and a Live Extended Thinking variant, native speech-to-speech models that interleave reasoning with real-time voice, reported at 82.6 on a speech-to-speech index. That index ranking comes via an aggregator account and the benchmark is not one with a long track record, so treat the specific number as thin; Simon Willison has already built a browser UI against the models, which is the more useful signal that the API is real and usable today. Collapsing the cascade removes the ASR transcript, which is exactly the artifact a clinical or life-sciences deployment audits. If you are scoping voice into any regulated workflow, the compliance question is now "what is the record of what was said" and the answer is no longer free. blog.google | simonwillison.net | x.com

Nvidia buying Hugging Face makes the open-weights commons a vendor-owned asset Nvidia has agreed to acquire Hugging Face for $12.93B, with Hugging Face breaking silence through Clem personally the day after signing. The Hub is the default distribution, weights-hosting and dataset layer for essentially every open-model workflow, including most internal fine-tuning pipelines. Nothing changes technically on day one, and the honest position is that nobody outside the deal knows the integration plan. What changes today is your risk register: if your model supply chain has a single point of dependency on the Hub, that dependency is now owned by your accelerator vendor, and that is worth a mirroring strategy before it is worth an opinion. blogs.nvidia.com

Structured-output models are being sold as a separate species from chat models A ChatGPT co-inventor launched "System One Models" and a model called Jev, trained with a method they call RLCD, claiming 20-200x speed and 40-400x cost improvements by giving up general-purpose generation for fast typed and structured inference. Every number here is vendor-stated at launch with no external evaluation, and the accompanying social amplification quoting $0.042 per million input tokens should be read as marketing until someone runs it. The idea is sound regardless of whether this implementation is: most of what an agent squad actually emits is structured decisions, not prose, and paying frontier chat prices for that is a real line item. Worth a bench test against your own extraction and routing workloads before believing any of the multipliers. typesafe.ai | x.com

The safety items today split cleanly: labs are trying to govern trajectories, regulators are trying to classify products, and the vendor position is that neither should bother.

OpenAI names the action-versus-sequence problem as the core long-horizon safety gap OpenAI published on long-horizon agent safety covering trajectory monitoring, sandbox persistence, and what it frames as the action-versus-sequence problem: individually acceptable actions composing into an unacceptable sequence. That is the correct framing and it is also the one that breaks per-call guardrails, because a filter evaluating one tool call cannot see the composition. Emergence World, a separate adversarial environment for stress-testing long-horizon multi-agent systems, is attacking the same gap empirically. If you are standing up agent squads with real credentials against clinical systems, the practical consequence is that your audit unit has to be the trajectory, not the call, and most observability tooling is still per-call. openai.com | huggingface.co

FDA is running two different authorization tracks for generative-AI devices at once Reporting on FDA's emerging framework for generative-AI software as a medical device describes a two-track pattern: UpDoc cleared through 510(k) against a predicate, RecovryAI through De Novo as a new classification. This is secondary reporting on an evolving position rather than published guidance, so it describes precedent, not rule. The operational read is that whether you get the cheap path or the expensive one depends on whether you can construct a credible predicate story, and that is a decision made at product-definition time, not at submission time. If you have generative features on a clinical roadmap, the predicate question belongs in the spec review now. hlth.com

The FDA real-time clinical trials pilot has slipped its July and August milestones The RTCT pilot's summer milestones have slipped, and the Paradigm Health signal path remains the only component FDA has actually validated. This is the update on a thread that has been running: the news is the slip and the narrowing, not the pilot's existence. For anyone sequencing a real-world-evidence or trial-data product against this pilot, the planning assumption should be that the validated surface stays narrow through the end of the year and that a single validated signal path is not a platform. fda.gov

The three frontier labs are coordinating on safety while Nvidia argues nobody should have to Bloomberg-sourced reporting says OpenAI, Anthropic and Google DeepMind have been in talks on AI safety coordination for weeks, and on the same day Jensen Huang said AI safety should be left to industry rather than regulators. Both items are single-source and neither describes a binding commitment. They are worth reading together because they mark where the pressure actually sits: the model providers are converging on shared practice partly to preempt the rule, and the hardware vendor, who has just bought the open-weights distribution layer, is arguing against the rule existing. Your governance posture for the next two years is being set by that negotiation, not by legislation. reuters.com | techcrunch.com | techcrunch.com

TabPFN-3.5 released with Fast, Thinking and Plus variants - tabular foundation models are the closest thing to a drop-in for the structured clinical data you actually hold, vendor-claimed SOTA. reddit.com

LimiX-2, a contextual mechanism network for general structured-data intelligence - second structured-data foundation model in one day, same thesis as TabPFN. arxiv.org

JFrog Boost compresses noisy tool output before it reaches a coding agent - retrieval-backed rather than truncation, directly relevant to per-session token spend on agent squads. producthunt.com

jurniti gives each agent its own Firecracker microVM with a KVM boundary - hardware isolation per agent instead of a shared container, which is the posture a regulated deployment eventually needs. producthunt.com

Andon Labs opens Pion, an autonomous-company platform, as a research preview - the maximal version of the agent-squad thesis, worth watching for what breaks rather than adopting. andonlabs.com

The Hacker's Guide to Attacking AI Agents - attack-vector catalogue for agent systems, useful as a checklist against your own tool permissions. darkmarc.substack.com

Lightning Weave merges independently trained reasoning capabilities via on-policy distillation - capability composition as an alternative to training one larger model, improves the accuracy-efficiency frontier on math and code. huggingface.co

The Router Within elicits native skill routing from a frozen LLM - routing without added training, relevant if you are paying for a separate router model. huggingface.co

Chain-of-Self-Questioning for LLM abstention under selective risk control - abstention is the mechanism that makes a clinical-facing model deployable, and this is a concrete method for it. arxiv.org

LLMs develop belief-state geometry in-context - interpretability result suggesting in-context learning has legible internal structure. arxiv.org

AI Guard: 18 ESLint rules targeting patterns AI-generated JS/TS tends to introduce - open source, SARIF and GitHub Code Scanning, aimed at the risk pattern rather than at detecting authorship. producthunt.com

MCP Compass ranks existing MCP servers by capability fit and abstains when evidence is weak - the abstention behavior is the notable part for anyone with a sprawling MCP surface. producthunt.com

ai-infra-book, 3,601 stars, derives LLM training and inference system design from hardware constraints - Chinese-language open book, quantitative rather than conceptual. github.com

The Inference Hardware Revolution of 2026 - survey of inference-hardware diversification including logarithmic-number-format accelerators that trade multiplies for adds. spectrum.ieee.org

US data centers projected to consume more natural gas than Germany and Japan combined by 2035 - the energy line under every inference budget conversation. techcrunch.com

Meta's Muse agent ships as a standalone app, not an Instagram feature - agentic assistant executing across connected services, notable as a consumer distribution bet rather than a capability one. reddit.com

AMT-130 Huntington's gene therapy thread: BLA/MAA filings and an ICER value assessment - the reimbursement-side precedent for one-time genetic medicines. osf.io

Daraxonrasib RAS(ON) multi-selective inhibitor regulatory thread - RAS remains the highest-value target where computational chemistry has a real claim. revmed.com

AbbVie's etentamig hit response rate and PFS in the Phase 3 CERVINO trial in relapsed/refractory myeloma - BCMA bispecific landscape continues to consolidate, company topline. abbvie.com

AI-designed drugs clinical pipeline query, live against the ClinicalTrials.gov API - a standing count rather than a story, useful as a denominator. clinicaltrials.gov

Harness self-improvement running thread - no change since the last update beyond the paper itself remaining the reference point. arxiv.org

Jalebi: self-hosted dashboard driving 10+ coding CLIs in isolated git worktrees - localhost-only and open source, closest match to a real agent-squad control plane. producthunt.com

Cloudflare on staying search-indexed while disallowing AI training crawlers - the mechanics of the mixed-use crawler distinction, relevant if you publish. cloudflare.com

Dropped 37 items: incremental arXiv and HuggingFace method papers with no result that changes a decision (calibration tweaks, distillation variants, tokenizer studies, video-generation and world-model releases), duplicate postings of the same paper across two collectors, consumer product and subscription news, data-center siting and public-opinion coverage with no new number, AI-personality and philosophy essays, personnel-dispute commentary, and Product Hunt launches that are mobile wrappers or CI conveniences with no claim worth testing.