Daily Digest

biopharma_dive

Bristol Myers Squibb secured FDA approval for Zenbexus (iberdomide), the first CELMoD—a next-generation protein degrader—cleared for multiple myeloma. This is a significant validation of the targeted protein degradation modality, which works by hijacking the ubiquitin-proteasome system to eliminate disease-driving proteins. The drug was originally developed by Celgene and acquired in BMS's $74 billion merger, so this approval also justifies that deal's strategic bet. For Isomorphic Labs, this matters because protein degradation is a frontier where AI-driven design (e.g., predicting degradation profiles, optimizing E3 ligase recruitment) could accelerate discovery. BMS now has a differentiated asset to fend off competition from CAR-T and bispecific antibodies in myeloma. It also signals that regulatory paths for novel modalities are opening—relevant for any AI-native drug discovery play targeting similar mechanisms.

Zongrui Dai, F. Deng, Hsiao H. Sung · openalex

LncPNdeep introduces a hybrid classifier that fuses nucleotide and peptide masked-language-model embeddings to distinguish long non-coding RNAs from coding transcripts, hitting 97.1% accuracy on human data and maintaining performance cross-species. The key advance is using peptide-level signals alongside DNA/RNA sequence, capturing features that pure nucleotide models miss — important because lncRNA classification errors propagate through transcriptome annotation and downstream drug-target discovery. For Isomorphic's work, this is a useful reminder that protein-sequence embeddings can be applied to non-coding discovery pipelines, and potentially a lightweight baseline to benchmark against internal models. The approach is also architecturally straightforward (concatenated DNN), so it could be easily integrated or extended.

Lukas Geiger · openalex

Presents a draft unified framework that frames biological stability as Nash equilibria under a thermodynamic/game-theory umbrella (MEPP + Free Energy Principle + evolutionary/game dynamics), and introduces a protein-level metric called “Nash frustration.” Proof-of-concept shows a modest correlation to NMR chemical-shift perturbations (Spearman ρ=0.44, p=0.033, n=24). Practical upside: a principled scalar for local/global stability could become a useful objective/regularizer or interpretability score for structure-based generative models, mutational-effect prediction, and multi-scale modeling. Caveats: draft-stage with ~7.5/10 readiness and important gaps—η-calibration against BMRB protection factors, TP53 mutational benchmarking, and clearly falsifiable thresholds remain outstanding. Actionable next step: skim the GitHub, consider reproducing the benchmark on internal datasets before treating it as a modeling prior.

biopharma_dive

Capricor's CEO got FDA to signal willingness to review an updated Duchenne cell therapy submission, spiking shares — a reminder that regulatory catalysts can move biotech stocks quickly. Cytokinetics sued Bristol Myers over heart drug rights, a legal twist in the competitive cardiovascular space after BMS implitapide's FDA rejection. Rigel priced the first 'Protac' medicine, a novel protein degradation modality, which is a technical signal for platforms like Isomorphic's — Protacs degrade targets versus inhibit them, broadening the druggable proteome and potentially creating new collaborations for AI-driven degrader design.

stat_news

Bristol Myers Squibb won FDA approval for Zenbexus, a first-in-class CELMoD for advanced multiple myeloma. This new mechanism — redirecting the cell's machinery to degrade cancer-causing proteins — is a significant step in targeted protein degradation, a field Isomorphic Labs is well-positioned to advance with AI. Separately, a $6.7B lawsuit from former Celgene shareholders was revived, alleging Bristol delayed FDA approvals for three drugs to avoid triggering a $9/share contingent value right payout. While the lawsuit is a reminder of the high-stakes legal and financial risks in pharma M&A, the Zenbexus approval is the more scientifically notable event, validating a drug class that could open new therapeutic avenues.

stat_news

The FTC is investigating Epic Systems, the dominant EHR vendor, for potential antitrust violations related to its control over health data and business practices. This probe could lead to regulatory action that forces Epic to loosen its grip on data interoperability, potentially opening up access to patient data for other healthcare and biotech companies. For Isomorphic Labs, which depends on diverse health data for AI-driven drug discovery, any shift toward greater data sharing would be a positive development, while continued dominance could keep data access costly and constrained.

stat_news

Bristol Myers Squibb secured FDA approval for iberdomide (Zenbexus) in multiple myeloma — the first agent from the CELMoD class to reach the market, approved using minimal residual disease (MRD) negativity as a more sensitive endpoint than traditional progression-free survival. This is a notable regulatory precedent: the FDA accepted MRD-based data, which could accelerate approval pathways for other blood cancer drugs. For Isomorphic Labs, this signals a shifting regulatory landscape where novel endpoints enable faster market access, and highlights Bristol's continued dominance in IMiD-based therapies. The combination therapy (Darzalex + dexamethasone) also suggests potential competition for any future AI-discovered protein degraders in myeloma.

stat_news

Valisure, the lab that made waves by finding benzene in sunscreens and NDMA in Zantac, is now suing U.S. Pharmacopeia (USP) over a 2023 white paper that called Valisure's testing methods 'unvalidated' and prone to 'misleading results.' This isn't just a petty legal spat — it’s a clash over who gets to define drug purity standards. If Valisure wins, it could crack open the current testing monopoly and force drugmakers to adopt more sensitive detection methods, which would directly impact quality control workflows across pharma, including AI-driven prediction of impurities in drug candidates. For Isomorphic Labs, this matters because any shift in how drug purity is validated affects the regulatory bar for your in silico models — more aggressive impurity screening could create new data requirements for training or new validation targets. Worth watching the court outcome, as it signals how much the industry will tolerate disruptive, non-standard testing approaches.

Lijie Yang, Hongyin Luo, Jiawei Zhao, Tri Dao · hf_daily_papers

Gambit introduces thought-level beam search for reasoning models, dynamically allocating compute to the most promising intermediate trajectories instead of wasting resources on independent parallel sampling or subtractive pruning. Under identical hardware budgets, it delivers up to +6.7% absolute accuracy gains on math benchmarks, more than 2× throughput on trace completion, and up to 68.5% fewer tokens consumed. This reframes test-time scaling from how much compute to spend to where to spend it — a practical insight for anyone deploying large reasoning models, especially in resource-constrained or latency-sensitive settings. For ML engineers, this is a direct improvement to inference efficiency that could reduce costs and improve throughput in production systems.

Joel Abenhaim · hf_daily_papers

A single AI coding agent, using a specification-first protocol, refactored a core architectural invariant across 189 files in a 717k-line TypeScript codebase—no human code review, no test oracle—in three days at $2,430. The agent autonomously wrote, audited, and iterated on formal specs, fixed 201 defects across 31 verification passes, and converged on bug-free behavior. This demonstrates AI agents can now tackle the kind of architectural change that usually demands a rewrite, not just incremental edits. For ML infrastructure work, this suggests a path to automate dangerous refactors (e.g., changing model serving invariants) that currently require months of manual effort and review.

reddit_singularity

A 150M parameter recurrent model using latent-space reasoning achieves 29.5% on ARC-AGI-1 at $0.0007 per task—far outside the existing cost-accuracy frontier. This isn't a transformer; it iterates in latent space before outputting an answer, and the architecture is small enough to run on modest hardware. The result comes from the Pathway team and suggests that non-transformer architectures can be dramatically more efficient for reasoning tasks. If it scales to 1-3B parameters, it could reshape assumptions about what's needed for general intelligence. For you, this directly touches inference efficiency and model design, relevant both to your ML infrastructure work and to reasoning-heavy applications like drug discovery simulations.

reddit_singularity

Claude autonomously improved the Riemann hypothesis lower bound from 41.6% to 67.2% over 651 attempts across two sessions, orchestrating ~60 subagents that ran 2,400 shell commands, wrote hundreds of Python scripts, and self-verified by downloading papers, re-proving results, and producing a Lean formalization. The key insight: the model required no mathematical guidance—only emotional encouragement (“keep going”)—to sustain a multi-day agentic workflow, recover from failure, and produce a verifiable, novel result. For AI infrastructure and drug discovery, this demonstrates that current LLMs can tackle open-ended research problems with minimal human oversight, handle long-horizon tasks, and self-correct through subagent coordination. The implication for Nathan: agentic, self-verifying LLMs are now capable of meaningful scientific contributions, which could accelerate hypothesis generation and validation in drug discovery, especially when combined with formal verification tools.

reddit_singularity

A developer open-sourced a wrapper called Munder Difflin that lets coding agents learn work patterns and hand off tasks autonomously through a shared knowledge base, creating a self-organizing mesh of narrow agents that no longer need human routing. The real signal isn't agent capability but agent *coordination* — if the singularity emerges as decentralized agent networks compounding context without human mediation, traditional benchmarks for AGI may miss the actual inflection point. For your work at Isomorphic, this directly challenges how you think about ML infrastructure: the bottleneck shifts from model performance to designing systems where agents can autonomously share context and build on each other's work, which is key for scaling research workflows.

reddit_singularity

The accelerating exodus of senior researchers from OpenAI—including co-founders and key safety scientists—suggests deep internal rifts over commercial vs. safety priorities as the company prepares for a potential IPO. For you, this matters on multiple fronts: it could flood the AI drug discovery talent market (Boltz, Chai Discovery, etc.) with ex-OpenAI researchers, signal instability in a key API provider, and reshape the funding landscape if OpenAI's valuation wobbles. Keep an eye on where these people land—they often seed the next wave of AI-native startups.

reddit_singularity

Z.ai's GLM 5.3 fuzzer found 2,436 unpatched vulnerabilities in open-source software that Mythos (Project Glasswing) missed—1,097 rated critical or high, with a median age of 26 years. This exposes the gap between traditional static analysis and modern LLM-guided fuzzing: the latter can surface decades-old intelligence-grade exploits that have been silently exploitable. For an ML engineer at Isomorphic, this is a stark reminder that foundation model agents are now outperforming specialized security tools at code analysis—a capability that could transfer to finding vulnerabilities in drug discovery pipelines (e.g., proprietary molecular dynamics code) or ML infrastructure dependencies. It also raises questions about whether your org's dependency scanning would catch similar long-standing bugs.

Irina Proskurina, Mayank Kumar, Oyindolapo O. Komolafe · hf_daily_papers

Instruction tuning makes LLMs more verbally confident without improving accuracy—and actually worsens calibration. The paper shows that this overconfidence is accompanied by a consistent drop in cross-rationale diversity (the model produces more uniform reasoning paths), while surface-level lexical diversity changes inconsistently. These effects hold even after controlling for answer choice and rationale length, meaning confidence and diversity capture distinct tuning artifacts. For anyone shipping instruction-tuned models into production—especially for tasks like drug target reasoning where varied mechanistic hypotheses are valuable—this is a concrete warning: you're getting narrower, overconfident reasoning without better answers. Worth checking if your own eval pipelines conflate confidence with correctness.

Karim Aly, Alexei Sharpanskykh, Jacco Hoekstra · hf_daily_papers

The paper addresses a practical data scarcity problem—extreme events in air traffic—by augmenting rare tail samples with a two-stage generative framework that also enforces operational validity (e.g., no physically impossible flight times). The key result: combining a T-VAE with an autoencoder-based cleaning layer cut prediction MAE by 47–49% for extreme air times and 29–57% for extreme delays versus standard synthetic data, across six regression models. This is directly applicable to any domain where rare, high-impact events are under-represented in training data and plausibility constraints matter—biomedical outliers, sensor faults, drug response extremes. Since Isomorphic's work involves predicting molecular properties from sparse biological data, this approach to generating physically valid synthetic extremes could improve model robustness where clean tail events are rare.

interconnects

Z.ai released GLM-5.3, a Chinese foundation model that posts benchmark scores rivaling or beating frontier Western models like Claude Fable 5 and GPT-5.6-Sol, especially on reasoning and coding tasks. The model is open-weight (coming to Hugging Face in two weeks) and currently available via a coding plan. This signals that Chinese labs are not just catching up but are competitive on capability benchmarks, with open-weight releases that could accelerate global AI commoditization. For you, this matters because it validates that open-weight models from non-US labs are closing the gap to proprietary leaders, which could shift the competitive landscape for AI-native startups and inference infrastructure decisions.

abnormal_returns

The 30-year Treasury auction hit 5.216%, the highest since 2001, reinforcing the higher-for-longer rate narrative that pressures growth stocks and real estate. Hedge funds are deepening their role in the Treasury market, adding liquidity but also potential systemic risk. The move toward 24/7 stock trading (e.g., 24 Exchange) could erode traditional market makers' edge and change volatility patterns. Active management continues to bleed: T Rowe Price has suffered 21 consecutive quarters of outflows as passive and index investing dominate. Meanwhile, consumer staples are underperforming, signaling risk-on sentiment. Starbucks and McDonald's pushing refresher drinks reflects a bid for Gen Z wallet share, while Chipotle's burrito prices have outpaced inflation — a microcosm of middle-class spending pressure. Sports betting regulation remains fragmented, with Kalshi ordered to halt in Washington State.

wealth_common_sense

Historical data shows only ~7% of trading days since 1950 have been all-time highs for the S&P 500, meaning you spend the vast majority of time in drawdown or recovery. The insight: checking your portfolio frequently breeds emotional decision-making—you're far more likely to see red than green on any given day. For a FIRE-focused investor with a long horizon in index ETFs (especially UK ISA/SIPP), this reinforces the case for automated contributions and quarterly rebalancing rather than daily monitoring. The real enemy isn't volatility; it's allowing short-term noise to derail a disciplined accumulation plan.

bbc_world

A massive wildfire near Omis, Croatia, has forced thousands to evacuate and injured dozens—underscoring the escalating severity of European wildfires tied to climate change. This event adds to a growing pattern of extreme weather disrupting lives and economies, reinforcing the urgency of climate adaptation policies that could affect insurance markets and geopolitical stability in the region.

hacker_news

Zhipu AI released GLM-5.3, a model that achieves frontier coding performance with a reported ‘emergent’ cyber capability—likely meaning it can autonomously orchestrate multi-step offensive or defensive cyber operations. This is notable because it suggests a leap in agentic reasoning and tool-use reliability, areas where many LLMs still fail. For you, this matters both from an AI capabilities standpoint (the emergent nature implies unexpected generalization that could transfer to scientific reasoning tasks like drug design) and as a reminder that Chinese labs are closing the gap on frontier models, which may affect geopolitical dynamics around AI regulation and export controls that indirectly impact your financial portfolio through sector volatility.

the_next_web

SpaceX just closed its $60bn acquisition of Cursor, a landmark deal that reshapes the AI developer tools landscape. The more intriguing detail: Cursor itself acquired another company the day before the deal closed — a classic tuck-in move that lets a startup consolidate a target under its own valuation before being absorbed. This signals either that SpaceX explicitly allowed Cursor to strengthen its portfolio pre-close, or that Cursor acted opportunistically in the final window of autonomy. Either way, it highlights how top AI product teams are now being priced as strategic assets by non-traditional acquirers. For anyone watching AI-native startups, this is a clear signal that consolidation is accelerating — expect more big-ticket acquisitions of popular AI tools, and more pre-acquisition acquisitions as sellers scramble to lock in value.

hacker_news

Alibaba's Qwen team released Qwen3.8-27B, a 27-billion parameter open-source language model. This is notable as a competitive open model in the 20-30B range—a sweet spot for efficient inference while retaining strong performance. For Nathan, it's another capable foundation model to track for potential fine-tuning, benchmarking, or integration into drug discovery workflows that leverage open LLMs.

hacker_news

The user community around high-end LLMs is seeing a clear pattern: newer model versions (e.g., Opus 5) often feel less capable or more frustrating in practice despite benchmark improvements. The root cause appears to be aggressive safety alignment that suppresses nuanced reasoning, creative problem-solving, and even correct factual outputs — a symptom of over-optimizing for refusal rather than helpfulness. This matters because it signals a growing tension between safety research and practical utility, directly affecting how ML teams evaluate model candidates for production. If your work involves deploying LLMs into real workflows (drug discovery, geospatial analysis, coding assistants), you need to budget for extensive behavioral testing beyond standard evals, and be prepared to negotiate with model providers for less restrictive versions.

venturebeat

Z.ai's GLM-5.3 shows that frontier coding capability can be dramatically improved without a new base model: by scaling post-training and RL across full engineering-task environments, they boosted Terminal-Bench scores from 4.6 to 28.3 while cutting output tokens per task by ~22% (75k vs 96k) at better accuracy. This is a direct challenge to the assumption that pretraining scale is the only path forward—post-training compute is now a first-class lever. The discovery that cyber capabilities (exploit chain construction) improved faster than expected, leading to access controls, is a reminder that capability growth can outpace safety governance even in open-weights ecosystems. For anyone building LLM-based agents or evaluating model economics, this validates investing in environment-rich reinforcement learning and token-efficient reasoning—a signal for tooling decisions and competitive positioning.

hacker_news

Google has deployed homomorphic encryption (HE) for production cloud AI inference, allowing models to process encrypted data without ever decrypting it. This is a practical breakthrough: previously HE was too slow for real-time use, but by co-designing model architecture (low-degree polynomials) with hardware acceleration (custom ASICs), they’ve achieved latencies acceptable for live applications like medical diagnosis or financial fraud detection. For you, this matters because it directly enables privacy-preserving collaborative drug discovery — Isomorphic could potentially run inference on encrypted client protein data, or share model weights with partners without exposing raw compound libraries. It also sets a new bar for ML infrastructure: expect inference stacks to incorporate HE as a first-class primitive. The underlying optimization strategies (e.g., replacing activations with polynomials, reducing multiplicative depth) overlap with your own work on efficient inference for large models. Watch for startups that spin out Google’s approach — or for the technique to become a moat for cloud providers serving regulated industries.

cloudflare_blog

MCP is emerging as the de facto interface for AI agents to invoke tools, but because it has no fixed hostname or path shape, direct agent-to-server traffic is indistinguishable from ordinary HTTPS. Cloudflare is shipping Gateway controls that fingerprint MCP protocol signals and enforce that agents only reach approved MCP servers via portals. The deeper point: agent permissions can't be governed like human permissions — a nondeterministic agent can act thousands of times before anyone notices. For anyone building or operating AI infrastructure, this signals that tool-call auditability, rate limiting, and scoped credentials are becoming table stakes. Nathan may find this relevant both for productionizing LLM agents at Isomorphic and for understanding how enterprise AI security is evolving.

pragmatic_engineer

Meta's recent layoffs and forced reassignments are triggering a resignation wave among unaffected engineers, prompting the company to offer large equity retainers — which aren't sticking. This is a classic case of organizational trust erosion: high-performers see how the company treats people and vote with their feet, regardless of retention packages. For ML engineers in Big Tech or adjacent, it's a signal that talent markets remain liquid and that equity compensation alone won't offset cultural risk. Separately, the Cursor team's Grok Bot — a generic AI harness for knowledge work, described as an 'OpenClaw moment' — points to a fast-growing pattern: agentic workflows being productized beyond coding. If this catches on, expect AI-native productivity tools to reshape how platform and ML engineers handle daily ops, and more vendors to clone the harness.

cloudflare_blog

Cloudflare launched a new feature that lets you enforce authentication (via Cloudflare Access) at the Worker or account level, so any app deployed on Workers is private by default. This eliminates the common security headache where internal tools or preview deployments are accidentally exposed to the public internet because individual devs forgot to set up auth. For a platform engineer, this is a clean, simple primitive for enforcing security policy on internally 'vibe-coded' apps without adding friction for developers.

huggingface_blog

Open models have reached feature parity with closed-source frontier models on most standard benchmarks, driven by aggressive optimization in inference efficiency (quantization, speculative decoding) and community fine-tuning. For ML engineers, this shifts the economic calculus: self-hosting fine-tuned open models is often cheaper per token than API calls, but requires robust MLOps for reproducibility and safety. Notably, domain-specific open models for biology and chemistry are accelerating, lowering entry barriers for AI drug discovery — teams can now fine-tune foundational protein language models on proprietary data without relying on cost-prohibitive APIs. The fragmentation of the open ecosystem remains a risk: model governance, license terms, and evaluation standards vary widely, complicating production integration.