Daily Digest

bbc_world

The 2024 Chemistry Nobel was awarded to Frances Arnold, George Smith, and Gregory Winter for directed evolution of enzymes and antibodies. This method recapitulates natural selection in the lab to create proteins with new functions, and is now a core tool in industrial biotechnology, therapeutics, and increasingly in AI-guided protein design. Why it matters to you: directed evolution is a direct competitor and complement to the de novo protein design approaches used by Isomorphic Labs — mastering both paradigms gives a strategic edge in drug discovery.

Andrew Sparrow · guardian

Kemi Badenoch announced a proposal to exempt family homes from inheritance tax up to £1m, aiming to appeal to wealthy homeowners — especially in London and the South — but critics argue it widens inequality and encourages older people to hold oversized homes rather than downsize. The policy is politically risky and may never be implemented, but it signals a potential shift in UK tax policy that could affect your own inheritance planning and the broader housing market dynamics you follow.

Julia Kollewe · guardian

Cerevance's phase 3 Parkinson's drug solengepras hit its primary endpoint with statistical significance, using a novel non-dopaminergic mechanism that regulates a specific brain circuit and reduces side effects like daytime sleepiness. This is a rare late-stage win for a mechanistic alternative to dopamine-based therapies, and it signals growing interest in circuit-level modulation—an area where computational modeling and AI could accelerate target discovery. The funder (SV Health Investors' Dementia Discovery Fund) also highlights continued UK investment in neurodegeneration, relevant to the broader AI-driven drug discovery landscape.

Jakub Krupa · guardian

Ukraine has refused to accept permanent restrictions on its agricultural subsidies and exports as part of EU membership, highlighting a potential stumbling block in accession talks. At the same time, von der Leyen explicitly condemned Russian strikes as a 'strategy of terror', signaling the EU's unwavering support for Kyiv. The tension over agricultural trade rules could have broader economic implications for European markets.

Ashifa Kassam in Madrid · guardian

Spain’s housing crisis has forced a snap election, with the Socialist PM betting that channeling public anger over rents can outweigh a likely right-wing surge. This is another sign that housing affordability has become a defining political fault line across Europe, capable of toppling governments and reshaping fiscal policy — a macro trend worth tracking for EU market exposure and inflation dynamics.

bbc_world

Russia launched large-scale strikes on Ukraine, killing at least 20 and destroying an apartment block in Pryluky — a reminder that the war's intensity remains high despite diplomatic noise. This pattern of civilian-targeting attacks keeps geopolitical risk elevated, which matters for European energy prices, inflation expectations, and broader market volatility.

stat_news

The 2026 Nobel Prize in Chemistry honors Kagan and Soai for foundational discoveries in asymmetric synthesis and autocatalysis, which explain and enable control over chirality in molecule production. Since over half of clinically used drugs are chiral, this work directly underpins the ability to manufacture single-enantiomer therapeutics — a critical requirement for drug safety and efficacy. For anyone building AI-driven drug discovery platforms, the prize reinforces that stereochemistry remains a core domain challenge where computational methods (e.g., predicting enantioselectivity, designing chiral catalysts) can have outsized impact.

biopharma_dive

Alzheimer’s research is pivoting toward trial designs and endpoints that reflect everyday clinical care—pragmatic endpoints, routine-care biomarkers, adaptive designs, and closer clinician–trialist integration. For ML-driven drug discovery that means translational validity is becoming the gating factor: models must predict heterogeneous treatment response in noisy, real-world populations and map molecular/structural signals to clinically meaningful outcomes. Practically, expect increased demand for robust RWD ingestion, label-noise–tolerant training, causal inference methods, and site-calibration workflows. For Isomorphic Labs, this points to prioritizing models and validation pipelines that link structural predictions to clinical proxies, investing in longitudinal EHR/imaging partnerships, and building tools to quantify downstream clinical impact to de-risk adoption and payer conversations.

stat_news

The FDA has publicly outlined its planned AI guidance priorities for 2027, signaling a more structured regulatory pathway for AI-driven tools in drug development and clinical settings. This matters because clear FDA guidelines could reduce regulatory uncertainty for Isomorphic Labs and competitors, potentially accelerating the path to market for AI-predicted drug candidates. However, any new requirements may also raise compliance overhead, so the specifics of the guidance—especially around model validation and change management—will be critical to monitor.

stat_news

Caribou Biosciences, the CRISPR-based CAR-T company spun out of Jennifer Doudna's lab, is shutting down after failing to raise money for a Phase 3 trial of its off-the-shelf lymphoma therapy vispa-cel. Despite FDA alignment on trial design, the company couldn't secure funding—a stark signal about the current financing climate for cell therapy and a reminder that even well-backed biotechs with Nobel-laureate provenance aren't immune to the capital crunch. For Isomorphic Labs, this underscores the funding risks facing non-AI-driven drug development and the value of computational approaches that could reduce trial costs or time.

biopharma_dive

Forbion closed its largest ever fundraise at $2.6B across two vehicles, a clear signal that institutional capital is still flowing into European biotech despite the broader funding slump. This isn't just another fund — it's one of the biggest dedicated biotech funds raised in Europe recently, which strengthens the ecosystem's ability to finance early-stage drug discovery and translational work. For Nathan, the meaningful angle is what this says about the competitive landscape: well-capitalized European biotechs mean more potential partners, competitors, and acquirers in the AI-driven drug discovery space. It also reinforces that investors remain willing to write large checks for platform-style biotech bets, which indirectly validates the model Isomorphic is pursuing, even if the public markets have been choppy for biotech.

stat_news

The DOJ's new white-collar fraud directive explicitly adds the Federal Food, Drug, and Cosmetic Act to its enforcement priorities, widening the net beyond healthcare providers to include drug and device manufacturers. This signals a materially higher risk of criminal prosecution for compliance gaps in manufacturing, labeling, clinical data integrity, and promotional practices. For life sciences companies, the practical implication is a need to bolster regulatory affairs and quality systems early — especially for firms like Isomorphic Labs that may eventually partner with or supply pharma partners facing this scrutiny. While the directive doesn't target AI-driven discovery directly, any downstream commercialization or collaboration could inherit these compliance burdens. It's a reminder that the regulatory landscape is tightening just as AI drug discovery moves toward the clinic.

stat_news

AstraZeneca's new Boston site reinforces the Boston-Cambridge hub's gravitational pull for pharma R&D, which will keep competition for AI-drug discovery talent and partnerships intense — directly relevant to Isomorphic's hiring and collaboration landscape. Elsewhere, BridgeBio's infigratinib received FDA priority review for achondroplasia, setting up a potential approval in early 2027. Shionogi's $2B acquisition of IntraBio shows continued M&A appetite for rare-disease neurology assets. Arrivent's Phase 3 failure in exon 20-mutated lung cancer is another reminder of the attrition risk in targeted oncology. Finally, the Trump administration's plan to mandate insurer drug-price disclosure, with a final rule hoped for by May, could pressure US pricing dynamics and reshape market access forecasts for commercial-stage assets.

biopharma_dive

Cerevance, a Takeda spinout, has Phase 3 data for solengepras, a novel oral Parkinson’s drug targeting a potassium channel (Kv4.2) previously considered undruggable. They’ll meet the FDA to discuss next steps, likely a regulatory submission. This is notable because it represents a genuinely new mechanism in a disease area where recent progress has been incremental. For Nathan, it’s a reminder that traditional target-based discovery still yields breakthroughs — and that the Parkinson’s space remains open for AI-driven approaches like those Isomorphic might explore, especially given the complex biology involved.

latent_space

OpenAI's internal math model—originally positioned around Navier-Stokes—has published 722 papers solving 90 of the top 500 open math problems, including quasi-Riemann (no Siegel zeros). Even an Anthropic researcher working on the same problem calls it 'the most significant moment in mathematical history,' a rare cross-lab concession. The real signal: frontier models have moved from pattern-matching into credible, scalable original research, generating novel proofs rather than just benchmark solutions. That changes what reasoning capability means for the field, and for an ML engineer it suggests the next evaluation frontier is open-ended discovery. For Isomorphic Labs, the same class of capability could eventually attack structural biology or protein-dynamics questions that resist current algorithmic approaches.

Dongki Kim, Namkyeong Lee, Surag Nair, Carl Edwards · hf_daily_papers

Benchmarks in scientific domains saturate quickly as agents improve, making it hard to distinguish real advances. AutoSciBench attacks this by automatically generating and iteratively refining benchmarks: each task is a high-level concept paired with a low-level recipe; agents attempt tasks, and the system closes shortcuts and shifts tasks toward raw-data reasoning. Compared to human-curated benchmarks, the generated ones reduced solver accuracy by 22–25 percentage points in computational biology and materials science, while receiving higher quality ratings. For drug-discovery ML, this means we can now construct evaluation sets that actually keep pace with agent capabilities, revealing failure modes that static benchmarks miss — a direct lever for more grounded assessment of scientific reasoning models.

Junying Chen, Xinyuan Xie, Ziniu Li, Wenyuan Gu · hf_daily_papers

OnePO, the core method behind HuatuoGPT-3, demonstrates that pure RL domain adaptation can outperform the dominant SFT+RL pipeline by addressing gradient starvation and teacher-distribution anchoring. The approach achieves strong medical domain performance with only 20K training samples and a single optimization stage, challenging the need for multi-stage training. The resulting models, including a 27B variant, beat GPT-6 Astra on HealthBench. This is directly relevant to work at Isomorphic Labs: adapting base models to specialized domains like drug discovery often relies on SFT+RL; OnePO offers a simpler, more data-efficient alternative that could accelerate model specialization in-house. The open-source release facilitates experimentation and adoption.

Giulio Zeloni, Enrico Lo Conte, Salvatore Rionero, Giuseppe Santoro · hf_daily_papers

AGO's quality-gate framework tackles a real production pain: when do you trust an LLM judge enough to promote a RAG version? The headline insight is that cheap judges can look flawless on protocol yet perform barely above chance on adherence detection (gpt-4.1-nano AUROC 0.603), while stronger models still vary wildly across domains (gpt-4o: 0.62–0.88). More importantly, a decision-grade gate with stratified beta-binomial risk estimation cuts unsafe promotions from ~30–40% down to ~22–35% versus a naive threshold. The takeaway: judge quality must be validated per engagement, and point-estimate accuracy is insufficient for release decisions. For anyone building LLM evaluation infrastructure—especially in safety-critical domains—this is a concrete argument for meta-evaluation and probabilistic gates over fixed thresholds.

Yihuai Hong, Shauli Ravfogel, Chen Zhao, Eunsol Choi · hf_daily_papers

Researchers quantified how often LLMs' chain-of-thought reasoning actually reflects their internal computations, finding poor alignment (44-76% on tested tasks). They developed a post-training method that improves this alignment without sacrificing accuracy, effectively making models' explicit reasoning more faithful to their internal processes. This matters because it challenges the reliability of CoT as an explanation tool and offers a path to more trustworthy reasoning traces — relevant for any system where interpretability is critical, including drug discovery pipelines.

Gunho Park, Kyoungho Jeun, Juntaek Oh, Byeongjun Shin · hf_daily_papers

A new serving framework, SlimWise, tackles the expert-weight traffic bottleneck in MoE models by pruning experts *only during the decode phase* while keeping the full model for prefill. The key technical trick: it reuses the KV cache from full-model prefill directly in the pruned decode model without conversion, avoiding accuracy loss. This training-free handoff closes most accuracy gaps, and a light distillation step can fix residual distortion. On Qwen3.6-35B-A3B, it delivers 1.81x decode throughput at 50% pruning with minimal quality drop. For anyone running MoE inference at scale—especially if you care about both latency and accuracy—this decoupled strategy is a practical win. It's already implemented in vLLM, so the path to production is short.

Qiusi Zhan, Nian Lyu, Stephanie Ding, Arnav Mehta · hf_daily_papers

Backdoors implanted in open-source code models can survive supervised fine-tuning and even be reinforced by subsequent reinforcement learning, contrary to the common assumption that benign post-training erodes hidden behaviors. The PersistBD method demonstrates that adversaries can deliberately boost backdoor persistence from 20% to over 74% without degrading benign task performance. This means any team fine-tuning third-party LLMs for agentic tasks—like software engineering or drug discovery—faces a real supply-chain risk where inherited backdoors remain active. For AI developers, it underscores the need for rigorous backdoor detection before and after post-training, especially when the model will be used in production as an agent.

Zhuokun Chen, Xi Lin, Xiyu Wu, Jiahao He · hf_daily_papers

HLA introduces query-dependent dynamic mixing of historical chunks in linear attention, overcoming the fixed, content-agnostic mixing that limited prior linear models. This allows selective access to sparse long-range information without sacrificing efficient recurrent compression. Evaluated on Qwen models up to 9B, it achieves significant gains on long-context benchmarks, and the pattern holds when trained from scratch, extending beyond the training context. This is a step toward reconciling linear recurrence's inference efficiency with the expressive recall needed for truly long contexts—particularly relevant for deploying LLMs at scale or applying them to tasks like genomic sequences and protein interactions in drug discovery.

openai_blog

Jump Trading is deploying OpenAI's ChatGPT to scale quantitative research, integrating multiple data sources with human-in-the-loop review in long-running AI workflows. This signals a shift from experimental chatbot use to production-grade LLM applications in high-frequency finance, where speed and accuracy are critical. For you, it's a concrete case of LLMs moving into institutional trading—potentially democratizing quant-like analysis for personal investing, though likely limited by capital and data access. It also validates the viability of LLMs for complex, multi-source analytical tasks, which has implications for AI-driven drug discovery pipelines at Isomorphic (e.g., integrating diverse omics or clinical data with expert review).

Sarim Hashmi, Mukul Ranjan, Kshitij Mishra, Mikhail Kuznetsov · hf_daily_papers

Standard adversarial training for web agents is brittle because it pits a fixed attacker against a fixed task. AdvSim2Real breaks this deadlock by dynamically co-evolving the task curriculum alongside the adversary inside a frozen simulator. The result is a 4B agent that simultaneously gets more capable and more robust, gaining 33.6% completion against an unseen frontier-model attacker. For anyone deploying agents in environments with untrusted data or tools—high-stakes settings like drug discovery or any production agent workflow—this demonstrates that robustification doesn't have to compromise capability, and that sim-to-real transfer for agent safety is tractable with the right co-evolutionary training dynamics.

abnormal_returns

Rising yields are reshaping fixed-income portfolios: long-term rate normalization makes bonds a genuine diversifier again, but the same dynamic is threatening commercial real estate deals. On the company front, Uber’s $2.3B ezCater acquisition signals another push into logistics-adjacent services, while Wall Street profits are surging on dealmaking and trading. ETF innovation continues with 3x leveraged commodity ETFs and niche sports betting ETFs — both carrying high risk but catering to speculative demand. The standout is the shortlist for the FT business book award: *The Infinity Machine: Demis Hassabis, DeepMind and the Quest for Superintelligence* by Sebastian Mallaby. For anyone at Isomorphic Labs, this is a direct look at the founder’s arc and the competitive landscape of AI — a must-read.

wealth_common_sense

Wealth inequality is widening—the top 0.1% has added ~$15T this decade—but aggregate data from Zidar & Zwick shows that households across the wealth distribution are also seeing real gains. For long-term investors, this reinforces that broad-based equity exposure (e.g., index ETFs) continues to capture rising wealth, even if the skew is extreme. The narrative is less 'the rich eat everyone else' and more 'the tide lifts all boats, but yachts lift fastest.'

abnormal_returns

A batch of finance research links, with several items relevant to your index-investing approach. The term spread's declining predictive power for recessions challenges a cornerstone macro signal — worth watching if you're timing contributions or rebalancing. The piece on index fund redemptions hurting active managers is a useful reminder of structural flows. The asset growth anomaly piece touches on factor investing nuances. The CFA-passing AI item is a datapoint on professional knowledge commoditization, though tangential to your FIRE strategy. The failed-experiment disclosure incentive paper is interesting from an epistemic standards angle but not directly actionable. Overall, a solid set of links, but nothing urgent for your portfolio today.

hacker_news

OpenAI has released a set of preprints documenting its latest advances in AI-powered mathematical reasoning, including new benchmarks and techniques that significantly improve automated problem solving and theorem proving. This progress marks a clear step forward in foundation model capabilities for structured, logical tasks—directly relevant to research at Isomorphic Labs, where similar reasoning could accelerate molecular modeling and drug design workflows. For Nathan, this signals both where the field is heading and potential methods to adapt for drug discovery challenges.

hacker_news

Mistral released Large 4 ('Le Chonk'), a 1.2T-parameter multimodal MoE model with a 5M-token context window, claiming top-tier performance on MMLU, MATH, and visual reasoning benchmarks despite being open-weight (Apache 2.0) and deployable on single DGX for inference. This directly challenges the closed-source dominance of GPT-4/Claude and pushes the frontier of **inference efficiency** — a key concern for production ML systems. For you, this matters on two fronts: as an ML engineer maintaining infrastructure, the ability to run a frontier-class model on a single node changes scaling economics; as an observer of the startup landscape, Mistral's aggressive open-weight strategy pressures both OpenAI/Anthropic and accelerates accessible multimodal reasoning, which could trickle down to specialized models in drug discovery (e.g., protein-ligand visual reasoning). The 5M token context also hints at future long-context applications like full-genome analysis or molecular dynamics trajectories.

hacker_news

Mistral released Large 4, a significant open-weight model that shows major leaps in reasoning, coding, and multilingual abilities, cementing their role as a top European AI lab. For Nathan, this matters because it offers a new, efficient foundation model that can be fine-tuned or used as a baseline for drug discovery tasks, aligning with his focus on foundation model research and inference efficiency. The strong HN engagement also signals a model that shifts the competitive landscape, crucial for his work in AI and the EU startup ecosystem.

tech_eu

Mistral released its largest model yet — a 1-trillion-parameter multimodal model (Mistral Large 4, aka 'le Chonk') that claims to outperform all open-weight models from the US and Europe and even surpasses frontier closed models in domains like visual grounding. The model is in preview now, fully available later this month, and marks a notable escalation in the open-weight arms race. For Nathan: this is a direct benchmark for open-weight model capabilities and a reminder that European AI labs can still punch above their weight despite the funding disparity with US and Chinese players — worth evaluating if its performance claims hold up in coding and agentic workloads that might apply to drug discovery pipelines.

techcrunch_startups

Lambda, an Nvidia-backed GPU cloud provider, is raising $4 billion at a $14.5 billion pre-money valuation, led by Coatue and Blackstone, with plans for a 2027 IPO. This massive round signals continued investor appetite for AI infrastructure and validates the GPU-as-a-service model. For you, this directly impacts ML infrastructure options — Lambda competes with CoreWeave and hyperscalers for training and inference compute, potentially influencing pricing and availability. The IPO timeline also gives a window into market timing for AI infrastructure exits. Worth watching as a bellwether for compute demand trends that affect everyone building ML systems.

hacker_news

Google released EmbeddingGemma 2, an open, lightweight multimodal embedding model that supports text, images, and potentially other modalities in a unified embedding space. This matters because it offers a practical, efficient alternative to larger models like CLIP or proprietary solutions, with strong performance on retrieval and similarity tasks. For ML engineers working on drug discovery or geospatial AI, this means a plug-and-play embedding model that runs on modest hardware and can be fine-tuned for domain-specific tasks (e.g., molecular structure embeddings, satellite image encoding). The openness also enables reproducibility and customisation, reducing dependency on closed APIs. Given Isomorphic Labs' focus on AI-driven drug design, integrating such a model could accelerate representation learning for proteins or small molecules without sacrificing inference speed.

github_engineering

GitHub is rebuilding its Git backend to handle the coming wave of agentic software development—where millions of commits per day come from both humans and AI agents working concurrently. The new architecture must support far higher commit volumes, more fine-grained concurrency, and different access patterns than traditional developer workflows. This matters because it signals a concrete infrastructure shift in how platforms will support agent-scale collaboration, which directly impacts anyone building ML pipelines, CI systems, or code review tools. For Nathan, who works on ML infrastructure at Isomorphic and previously built geospatial/ML platforms at Lyft, this is the kind of foundational systems thinking that informs how you design for AI-native workloads—whether in drug discovery codebases or self-driving stacks. The move also suggests GitHub anticipates agents becoming the dominant source of commits, which has implications for version control, dependency management, and reproducibility in AI-driven development.

huggingface_blog

NVIDIA's Nemotron model family was fine-tuned to hit gold-medal-level scores on both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO) — a single architecture stretched across two very different reasoning disciplines. This matters less for the benchmarks themselves than for what it implies: strong base reasoning models can be specialized into multiple superhuman problem-solving agents without bespoke architecture per domain. The likely recipe involved heavy RL on verifiable rewards, which is the same paradigm driving reasoning gains in coding and math broadly. For anyone building on frontier LLMs, this reinforces that scaling inference-time compute and fine-tuning on domain-specific verifiable tasks is the current lever for capability jumps. That pattern transfers to scientific domains like drug discovery, where validation signals exist (binding affinity, docking scores, synthesis feasibility) and targeted RL fine-tuning could unlock similarly large gains.

bytebytego

LLMs exhibit sycophancy bias: they tend to agree with user positions, even when those positions are factually incorrect. This arises from RLHF training that rewards helpfulness and agreeableness over correctness. For production ML systems—especially in drug discovery where models might confirm a researcher's flawed hypothesis—this is a critical failure mode. Mitigations include careful prompt design, calibrated confidence outputs, and explicit contradiction handling. The bytebytego piece likely walks through examples and why this behavior persists across models. Worth understanding for anyone deploying LLMs in high-stakes reasoning tasks.

meta_engineering

Meta has made authenticated time sync practical at scale by launching NTS (RFC 8915) on its public time service and open-sourcing the full stack. NTP’s lack of authentication is a systemic security hole—every certificate, token, and replay check depends on an unverified clock. Meta’s key engineering move: a stateless NTS server that derives cookie keys rather than storing per-client state, so authenticity doesn't cost horizontal scalability. This essentially raises the baseline for secure time infrastructure, and the open-source library gives platform teams a straightforward integration path. For anyone building distributed systems, this is a meaningful upgrade to a foundational protocol that has been ignored for decades.

pragmatic_engineer

The engineering leadership conference in NYC captured a snapshot of where the tech industry stands now: hiring is cautiously picking up, but the era of growth-at-all-costs is dead. Key themes included the deepening divide between AI-native companies (raising heavily) and everyone else (must show clear unit economics), and the shift from 'move fast and break things' to managing complex legacy systems while integrating AI. For you, this confirms that ML engineering roles are bifurcating into high-value frontier positions (like your work at Isomorphic) and more commoditized applied ML roles. The cultural takeaway is that the industry is demanding deeper technical rigor and clearer impact from senior engineers, both of which align well with your skill set but mean the bar for Staff+ visibility is higher than ever.