Daily Digest

stat_news

Roche is shifting its China strategy from licensing mature assets to early-stage discovery by partnering with Defand Therapeutics, a Chinese biotech startup. This signals growing confidence in Chinese discovery capabilities and could accelerate deal models that value novel platform technologies—relevant for AI-driven drug discovery companies like Isomorphic Labs that seek similar partnerships. The move also reflects broader pharma interest in accessing diverse preclinical pipelines, potentially increasing competition for talent and targets in the AI biotech space.

Fleming Kretschmer, Eva-Maria Harrieder, Michael A. Witting, Sebastian Böcker · openalex

A new two-step ML approach finally solves a long-standing pain point in small-molecule LC-MS analysis: predicting retention times on chromatographic systems the model has never seen. By first predicting a retention order index conditioned on chromatographic conditions, then mapping that index to absolute times, the model decouples order from scale—something previous models conflated. It beats models trained directly on the target system, meaning labs can skip costly calibration runs. For drug discovery, this is directly applicable to metabolite identification, impurity profiling, and untargeted metabolomics where reference standards are unavailable. Given your ML-for-chemistry background at Isomorphic, this is a clean example of inductive bias (order invariance) beating brute-force training, and it hints at broader lessons for transferable molecular property prediction.

stat_news

The Trump administration is probing nine universities for J-1 visa fraud targeting postdoctoral researchers. This is a direct threat to the biomedical talent pipeline — postdocs are the backbone of academic and early-stage drug discovery. For Isomorphic Labs and competitors, tighter visa enforcement could slow recruitment of international researchers, especially in AI-driven structural biology where many top candidates come from these universities. Watch for spillover effects on industry partnerships and academic collaborations.

stat_news

NIH funding uncertainty is tightening, casting a pall over U.S. biomedical researchers and spooking biotech investors. The latest funding cycle distribution reveals growing unpredictability, which could slow early-stage research pipelines and make it harder to recruit or collaborate with top academic labs. For Isomorphic Labs, this means a potential contraction in the U.S. research talent pool and fewer academic partnerships, though it may also drive more talent toward well-funded AI-native startups like yours.

stat_news

The newsletter highlights three health tech developments: AI for breast cancer risk prediction moving direct-to-consumer, raising questions about clinical oversight and patient interpretation; a regulatory 'sandbox' in Utah facing scrutiny that could set precedents for AI oversight in healthcare; and the emerging phenomenon of AI 'psychosis' — likely referring to model hallucination or unreliability — underscoring the challenge of deploying LLMs in clinical settings. For Nathan, the DTC AI trend could influence how Isomorphic's drug discovery models are eventually used, while regulatory sandboxes and AI reliability are directly relevant to his work in AI-driven pharma.

Jia Zhang, Siqi He, Qiguo Xiao, Weijie Jiang · openalex

A new study demonstrates that engineered stem cell-derived extracellular vesicles overexpressing miR-146a can reduce neuroinflammation in Alzheimer's models by downregulating NR4A3 and suppressing NF-κB signaling, leading to improved cognitive function. This positions miR-146a and the NR4A3–NF-κB axis as a novel therapeutic target for neurodegenerative diseases, and vesicle-based miRNA delivery as a viable modality. For AI-driven drug discovery, understanding such regulatory mechanisms could inform target prioritization and modality selection beyond traditional small molecules.

stat_news

The 2030 U.S. census methodology changes will artificially inflate mortality and morbidity rates by shrinking the population denominators used in health statistics. The effect will be uneven across geographies and demographic groups—especially college-age and retirement-age people, people of color, and immigrants. That means public health funding, disease surveillance, and even real-world evidence used in pharma analytics could be systematically skewed for years. For anyone relying on U.S. health data to size markets, design clinical trials, or assess disease burden, this is a hidden data-quality landmine. It also highlights why census investments are actually public-health infrastructure, not just politics.

stat_news

With Democrats favored to retake the House in the midterms, a wave of new lawmakers is competing for seats on the Energy and Commerce and Ways and Means committees — both with broad jurisdiction over drug pricing, FDA reform, and Medicare/Medicaid policy. This could shift the balance of power on health care oversight from pro-industry incumbents to progressives pushing for aggressive drug price negotiation and tighter regulation of pharma. For AI drug discovery, that means a more volatile policy environment: potential pressure on pricing models and intellectual property could alter how Isomorphic and peers monetize computational drug candidates, while FDA modernization efforts might either accelerate or complicate approval pathways. Worth monitoring which fresh faces land the gavels that set the health care agenda starting in 2027.

Pengxiang Li, Dilxat Muhtar, Di He, Guinan Su · hf_daily_papers

Looped transformers degrade with more iterations because noisy updates overwrite correct intermediate states. InfiLoop solves this with a residual connection that learns a running summary of past computations using content-based weighting and temporal decay, keeping a constant memory footprint. The result is a 7M parameter model that improves with test-time looping beyond 20,000 steps, reaching 97.9% on Sudoku-Extreme and 13.6% on ARC-AGI-2. This demonstrates that recurrence in transformers can actually benefit from massive test-time compute if the state updates are adaptively filtered — a practical insight for any production system where reasoning depth matters.

Tobias Braun, Nils Loose, Alexander Herzog, Virginia Ceccatelli · hf_daily_papers

U-Space introduces a method to quantify uncertainty in LLMs by constructing a low-dimensional subspace from semantic anchors for doubt and certainty, then projecting token states onto it. This yields interpretable token-level uncertainty maps without requiring correctness labels, repeated generations, or training. The approach outperforms existing baselines on reasoning benchmarks even when controlling for output length, indicating true uncertainty signal rather than length artifacts. For Nathan, this directly addresses a critical pain point: deploying LLMs in drug discovery where incorrect confident outputs have high cost. The method's interpretability (localising uncertainty to specific tokens) and zero-shot nature make it practical for production systems, and the mechanistic interpretability lens aligns with current work on understanding model internals.

Jitai Hao, Quansheng Gu, Qiang Huang, Jun Yu · hf_daily_papers

SparseEngine introduces a sparse-first inference engine with a shared lifecycle contract that allows flexible KV representation and computation control. It achieves over 10x higher throughput with KV eviction and over 2.5x faster decoding than vLLM at matched concurrency, while also enabling cross-request state management via Chain Cache and Prefix-Cache Pruning. This matters because it directly addresses the bottleneck of long-context LLM agents, offering substantial performance improvements for serving models at scale—relevant for any production ML system, including potential drug discovery applications that rely on LLMs for molecule understanding.

Shuyu Gan, Young-Jun Lee, Dongyeop Kang · hf_daily_papers

SanSi introduces a 'System 1.5' thinking mode for language models: recursively applying the same layers multiple times before a single typed readout, without generating any tokens. This looped architecture reaches 72% accuracy on 10k decision tasks—13.5 points above a non-looped model of identical size and only 1.8 points below a model with 3x the parameters. The loops also extend the solvable reasoning depth beyond what the model saw during training, where the larger static model fails. As an RL judge, it improved generator F1 by 7.7 points without gold answers. For ML engineers, this means a cheap way to buy significant capability gains at inference time: you can dynamically choose the number of loops (budget) without retraining, and the approach generalizes to deeper reasoning. This could be especially valuable for fast, accurate predictions in drug discovery (e.g., binding affinity classification) where you want more than a single forward pass but can't afford full token generation.

openai_blog

LegalOn cut estimated daily Codex costs by 65% while maintaining development speed by routing tasks among multiple models—Astra, Sol, and Luna—and strategically managing budgets. This is a concrete example of how intelligent model assignment can slash inference costs without sacrificing quality. For Nathan at Isomorphic Labs, where LLM inference costs in drug discovery pipelines are a real concern, this demonstrates that significant cost reduction is achievable without slowing down. It also speaks to his broader interest in ML infrastructure and platform engineering: the ability to design systems that dynamically allocate workloads to the most cost-effective model tier is a valuable capability. Beyond the specifics, it highlights a trend toward more efficient, multi-model architectures across production AI systems.

latent_space

Periodic Labs, founded last September by Liam Fedus (ex-OpenAI) and Ekin Dogus Cubuk (ex-DeepMind), is rapidly emerging as a dominant force in AI-driven scientific discovery. With a founding team that includes co-creators of ChatGPT, DeepMind’s GNoME, and the neural attention mechanism, they are building industrial-scale autonomous physics labs to accelerate science. Their talent density and progress pose a direct competitive threat to Isomorphic Labs in the AI scientist space, especially for drug discovery applications.

reddit_ml

The ThinkingBox benchmark reveals that agent reliability is starkly different from discovery capability: Kimi-K3 solves 93.9% of tasks at least once but only 13.4% every time across 20 trials, while Claude Opus 5 discovers fewer (79.1%) but repeats far more (47.5%). Over 67% of failures terminate cleanly but produce wrong database state — a reminder that terminal success metrics (pass@20) can mask severe unreliability (all-20). For any production system where agents touch real data, all-20 matters more than pass@k, and clean-looking failures are the norm, not the exception.

Tianyu Fu, Tengxuan Liu, Ruoxi Wang, Yixin Dong · hf_daily_papers

Token-level LLM routing has been stuck in theory because existing serving systems are built around single-LLM assumptions, causing step desynchronization and batch admission delays. TokenRouter fixes this by decoupling the request view from execution: developers write routing logic per-request while the runtime launches separate subservers per model and dispatches asynchronously. Its delayed-batching scheduler is tuned from a throughput model, and the result is 2–64x higher decoding throughput across routing algorithms and workloads. This matters because it makes fine-grained routing economically viable in production—a step toward treating model choice as a per-token, cost-quality tradeoff rather than a static deployment decision. For anyone running multi-model inference stacks, that changes what serving infrastructure should look like: routing isn’t a front-end concern but a core scheduler primitive.

Rongxue Li, Meng Yang, Yiru Mao, Yongliang Tao · hf_daily_papers

A new self-distillation framework, SpatialOPSD, enables multimodal LLMs to internalize spatial reasoning without relying on external coding agents at inference time. The key insight is that prompting an MLLM with summarized execution traces from a spatial coding agent unlocks internal chain-of-thought reasoning, which can then be distilled back into the model using a repetition-aware unlikelihood regularization to prevent privilege leakage. The result: higher accuracy and better generalization than standard SFT or GRPO on both spatial and out-of-distribution benchmarks. For you, this is directly relevant to geospatial AI—it suggests a path to more efficient, tool-free spatial reasoning in production systems, and the distillation technique itself is noteworthy for anyone working on model compression or capability internalization.

Core Team, Zongming Qiao, Ziyue Hua, Zirui Ou · hf_daily_papers

MiMo-V2.6 shows that scaling reinforcement learning across batch size, environment diversity, and grader compute enables self-improvement in omni-modal models. Key engineering choices include hybrid-SWA pretraining, freezing MoE routers for stability, and groupwise agentic grading to produce token-efficient solutions. The paper also details infrastructure for mixed-task agentic RL with high-concurrency rollout and decoupled control/data planes. This is directly relevant to training large models at scale—expect to see these patterns adopted in foundation model pipelines, including those for drug discovery. The open-source release of training dynamics and RL environments makes it easy to adapt.

Adam Morton in Nadi, Fiji · guardian

The Cop31 president-designate is reframing climate finance as existential security, pushing wealthy nations to outspend their 5% Nato defence commitments on climate mitigation — a signal that the climate-security nexus is becoming central to summit diplomacy. His hedging on repeating the fossil fuel transition pledge suggests Turkey may soften the language at Antalya, potentially weakening the strongest commitment from previous COPs. For Nathan, this is a bellwether for how geopolitical instability and US climate denial are shaping international climate policy heading into COP31.

Alexandra Villarreal in New York · guardian

Trump's immigration crackdown has driven out 1.2 million foreign-born workers, reversing prior price trends: immigrant-dependent foods now inflate faster, new-home costs are up ~11% with permits down ~11%, and by 2034 Social Security benefits could shrink by ~$2,150/year as the labor-supply shock erodes economic growth and payroll-tax funding. The data undercuts the administration's claims and signals persistent inflationary pressure and slower US growth — worth factoring into macro views and long-term portfolio assumptions.

bbc_world

OpenAI has fired researchers who claim they were let go for prioritizing safety; the company counters they mishandled sensitive data. This incident underscores the ongoing tension between safety culture and commercial pressures at frontier labs, a dynamic that shapes talent movement and regulatory debates in AI—relevant to anyone tracking alignment progress and organizational trust in the field.

bbc_world

The USS Lincoln's extended 10-month deployment, driven by the war with Iran, signals a protracted US military commitment in the Middle East and sustained regional instability. This directly impacts global energy markets and supply chain resilience, offering a macro data point for assessing geopolitical risks to personal portfolios.

Kalyeena Makortoff · guardian

China and the EU reached a 16-point agreement to de-escalate trade tensions, exchanging relaxed rare earth export restrictions for concessions on electric vehicle exports — a notable shift in the geopolitical trade landscape. Meanwhile, UK diesel has dipped below £2/litre but remains 57p above pre-US/Iran conflict levels, with scheduled fuel duty increases looming; the government faces pressure to provide relief in the upcoming budget. Separately, the UK Court of Appeal overturned another Libor trader's conviction, further unraveling the SFO's landmark prosecution.

Martin Belam (now) and Jakub Krupa (earlier) · guardian

The Nobel Peace Prize awarded to Navi Pillay—who led a UN report describing Gaza as genocide—has drawn sharp condemnation from Israel, highlighting the growing clash between international legal institutions and powerful states that reject their authority. This signals a broader geopolitical shift where rule-of-law mechanisms are increasingly weaponized or dismissed, which could affect global stability and institutional trust.

monevator

Replicating the UK State Pension's inflation-linked, lifetime income (currently ~£12.5k/year) via an RPI-linked annuity costs around £225k at age 65, based on Hargreaves Lansdown best-buy rates (£5,566 per £100k). That's a sobering figure for anyone betting the triple lock won't survive — it underscores how valuable the State Pension is as a baseline for FIRE planning, and why even modest reductions in its generosity would be costly to DIY.

wealth_common_sense

With bond yields around 5%, the stock-versus-bonds question is no longer one-sided: fixed income now offers a credible, low-risk return stream that can fund retirement spending without relying on equity volatility. For a FIRE-focused investor, the real insight is that locking in ~5% yields lets you prioritize cash-flow certainty over total-return maximization, especially if your withdrawal rate is already covered. The piece also flags rising debt risks among hyperscalers — relevant if your index exposure is heavily concentrated in mega-cap tech, since those balance sheets underpin much of recent equity returns. It's a useful reminder to stress-test portfolio assumptions beyond simple historical averages.

abnormal_returns

The bond sell-off is partly technical (mortgage convexity unwind) but signals real rate concerns. Credit markets show stress: SpaceX CDS rising, private credit valuation issues. Meanwhile, tokenized US stocks (Securitize) launch on Solana, hinting at new liquid asset classes. For your portfolio, monitor bond ETF exposure—the convexity unwind may reverse—and watch credit spreads for systemic risk signals. Tokenized assets could offer tax-efficient access (ISA/SIPP?) but regulatory clarity is needed.

abnormal_returns

Prediction markets are maturing into a serious asset class, as elite traders systematically clean up on Polymarket — this mirrors the broader financialization of gambling and has implications for portfolio hedging and alternative data. The Larry Ellison and Nubank profiles reinforce how tech founders build financial infrastructure (data monopolies, neobanking) that directly interacts with retail investing. Geopolitical and demographic risks (China's Fed infiltration, fertility traps) are also surfaced, which macro-focused portfolios should track. The rest of the links (cognition essays, climate PR, personal narratives) are less actionable for a finance lens, but the gambling pieces are standout for anyone monitoring market structure changes.

hacker_news

OpenAI's annualized revenue has landed $20B below what earlier signals suggested, a rare concrete data point that separates hype from actual AI monetization. This matters beyond OpenAI: it resets the benchmark for what frontier labs are worth and pressures the whole AI infrastructure stack — Nvidia, Oracle, CoreWeave — whose valuations assume exponential demand growth. For AI-native startups, this means tighter enterprise budgets and a harder fundraising environment as investors re-baseline revenue multiples. It also strengthens the case for domain-specific AI plays like drug discovery, where measurable value creation is clearer than general assistant subscriptions. Watch for downstream effects on compute pricing and partnership structures in the next few quarters.

hacker_news

4-hour battery storage has reached cost parity with—and is now cheaper to install than—gas turbines on a global scale. This flips the economic logic of peaker plants, which were the last stronghold for fossil fuels in electricity grids. For your portfolio, this signals that renewables + storage are no longer a niche substitute but the default new-build option, accelerating coal and gas retirement timelines. For startup watchers, expect a wave of grid-scale battery companies, virtual power plant models, and software to optimize dispatch. The macro implication: faster decarbonization than most models assume, which will reshape energy commodity prices and utility business models globally.

hacker_news

Cactus Compute released Whistle, a speech-to-text model compressed into just 16.9 MB — roughly 200x smaller than typical Whisper variants while retaining competitive accuracy. This is a concrete demonstration that aggressive distillation and architecture search can collapse model size without catastrophic quality loss, opening the door to fully on-device transcription on microcontrollers and edge hardware. For Nathan, the significance isn't the STT use case itself but the proof that extreme compression is viable in production: it validates a frontier in inference efficiency that carries over to any latency-sensitive or resource-constrained ML deployment, including geospatial edge processing and possibly lightweight predictive models in drug discovery pipelines.

sifted

The UK is set to outlaw non-compete clauses, a move that directly reshapes talent dynamics in the startup ecosystem. For founders and early-stage companies, this removes a defensive tool, making it easier for employees to leave for competitors or spin out new ventures — which historically has been a significant drag on UK innovation. The practical effect is increased labor mobility and knowledge transfer, particularly in high-skill sectors like AI and biotech. For you personally as a London-based ML engineer, this is double-edged: it strengthens your own career optionality and startup opportunities, but it also means Isomorphic Labs has one less legal lever to retain key talent in a fiercely competitive hiring market. Early-stage UK startups will benefit most, as they can now poach senior talent from incumbents without legal friction, potentially accelerating the pace of new AI-native company formation.

crunchbase_news

Europe's Q3 2026 venture funding hit $25B — its strongest quarter in four years — with AI companies accounting for 75% of that total ($18.8B). Mistral's $3.5B round (the largest European venture round ever) and big defense tech raises (Helsing, Quantum Systems) dominated, but the broader signal is that Germany and France had their best quarters since 2021, showing the ecosystem is finally diversifying beyond the UK. For someone tracking the EU/UK startup scene and AI-native companies, this confirms that capital is flowing into European AI and physical tech at near-boom levels, which could benefit spinouts, talent mobility, and early-stage deal flow in London, Paris, and Berlin.

tech_eu

The UK government will legislate to ban non-compete clauses for startups and scaleups, a move that directly targets one of the biggest friction points in the UK tech talent market. Inspired by campaigns from ElevenLabs, Synthesia, and the Startup Coalition, the policy is designed to accelerate hiring and company-building by removing barriers that prevent employees from moving to or founding competing ventures. For you, this is a structural shift in the UK startup ecosystem: it lowers the risk of joining an early-stage company, increases talent fluidity, and makes London more competitive with the Valley for top ML and engineering talent. Even if you're not currently job-hunting, it signals a healthier environment for spinouts and AI-native startups—which could affect future opportunities and the broader talent pool at places like Isomorphic Labs.

huggingface_blog

The path to solving an ML problem increasingly runs through building your own model rather than waiting for one to appear. With open weights, efficient fine-tuning, and accessible tooling, the marginal cost of a bespoke model has dropped to near zero—so the default question shifts from 'is there a model for this?' to 'what's the cheapest way to train or adapt one?' For someone shipping AI products, that changes how you evaluate feasibility: custom models are no longer a moonshot but a routine engineering option. It also means the bottleneck moves to data and evaluation, not architecture. The practical takeaway: invest in the pipeline around the model—data curation, eval harnesses, and iteration speed—because that's where the real leverage lives now.

pragmatic_engineer

Ramp and Stripe have both built internal platforms that let non-engineers create their own websites and tools without writing production code — this is a shift in how internal tooling gets built at mid-sized tech companies. The platforms are gaining traction inside those companies, and this pattern is likely to spread. For someone with your ML platform background, this signals a growing expectation for internal platforms to support rapid, safe self-serve tooling, which could affect how you design or advocate for internal infrastructure at Isomorphic Labs.

bytebytego

Defining software boundaries—deciding which responsibilities live together and how components communicate—is a more impactful skill than raw coding ability. Poor boundaries lead to tangled dependencies, slow iteration, and brittle systems; good ones preserve modularity and enable independent scaling or replacement of parts. For someone building ML infrastructure or platform teams, this concept directly applies to separating model training, inference, data pipelines, and experiment tracking. A reminder that the most important design decisions aren't about algorithms or frameworks, but about where you draw the lines.

cloudflare_blog

Cloudflare redesigned Radar to make its global internet data more accessible to non-technical users (journalists, policymakers) without sacrificing depth for network operators and researchers. The new design uses a map-based landing page and tabbed navigation to guide users from high-level trends into detailed analysis, addressing a common tension in data platforms. For Nathan, this is a case study in balancing technical rigor with usability—relevant to ML infrastructure and platform engineering where similar tradeoffs arise.

huggingface_blog

Falcon ASR, a new open-source automatic speech recognition model, has been released on Hugging Face. It extends the Falcon family into audio, likely offering strong performance with efficient inference. This matters as another high-quality foundation model available for fine-tuning or integration, reflecting ongoing investment in open-source AI and potentially enabling new multimodal applications.