Daily Digest

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

Russian drone strike deliberately targeted Ukraine's National Academy of Sciences in Kyiv, killing one and wounding several, in what the director termed an 'attack on science.' This escalation underscores the growing threat to civilian research infrastructure and academia during the war. Poland's concurrent decision to authorize shooting down foreign aircraft based on radar signals a significant hardening of Nato's eastern flank.

Sara Sneath · guardian

A new study traces how decades of libertarian donor networks, fossil fuel companies, and corporate-funded academic centers systematically eroded US university autonomy, making institutions highly vulnerable to the Trump administration's 2025 funding freezes and political pressure. The Academic Freedom Index shows US institutional autonomy collapsed from 3.3 in 2019 to 1.7 in 2025. For scientists, self-censorship is now structural: grant applications avoiding words like 'climate change', programs preemptively shuttered, and researchers weighing whether a question is worth losing funding. This isn't just a US political story — it signals a chilling effect on the entire evidence base for climate policy, and a blueprint for how corporate interests capture research agendas globally, with direct parallels to how industry funding shapes AI and biomedical research incentives.

Shrai Popat (now) and Tom Ambrose (earlier) · guardian

Iran publicly rebuffs Trump's claims of imminent US negotiations, exposing a gap between White House rhetoric and diplomatic reality. Domestically, Trump's midterm push leans on 'pretend I'm on the ballot' appeals while his 'greatest economy' boast collides with 3.4% inflation, soaring gas prices, and a Fed rate hike; his Xi summit and private dinner with Anthropic's Dario Amodei signal AI is now a geopolitical lever. The economic data matters for anyone watching market stability, and the AI-policy courtship is directly relevant to Nathan's field.

Lauren Almeida · guardian

Nvidia's $150bn buyback—the largest ever by a US company—signals extraordinary cash flow from AI chip sales and confidence in continued AI growth. It underscores the scale of the AI investment wave and Nvidia's central role.

Kat Lay · guardian

The US blocked a UN political declaration on pandemic preparedness, signaling a retreat from multilateral health cooperation and weakening global planning. This move, aligned with Trump administration's broader stance against international institutions, raises concerns about future collaboration on pathogen sharing and equitable access to treatments—issues that directly affect the environment for global drug discovery partnerships.

bbc_world

The war in Yemen is escalating, with pro-government forces reporting from the front-line indicating sustained conflict. This prolongs instability near the Bab el-Mandeb strait, a key chokepoint for global oil shipments, which could introduce volatility in energy prices and shipping costs—factors that ripple into global inflation and affect the macro backdrop for portfolio management.

Yan Zhan, Shaobo Liu, Qiunan Liu, Yuanjun Shi · hf_daily_papers

Tool-calling RL suffers a subtle but serious flaw: standard GRPO assigns one trajectory-level advantage to every token, so noise from natural-language summary tokens contaminates the tool-decision tokens, producing brittle agents. SLCA-GRPO fixes this with segment-locked credit assignment, computing separate advantages for tool and summary segments within the same rollout, eliminating the contamination without needing extra rollouts. It also introduces a schema-guided LLM simulator to replace expensive real API calls during training. On a 7B backbone, this converges faster and beats GRPO, ToolPO, and RLTR by up to +9.15 pp on τ²-Bench while cutting tool redundancy. For anyone training agentic models—especially in drug discovery where agents must call external tools cleanly—this is a practical, likely transferable optimization improvement.

Chenyu Su, Zhaolong Shen, Yuan Qian, Chen Qian · hf_daily_papers

VLA-Precision tackles the problem that pretrained vision-language-action models fail at high-precision manipulation tasks (e.g., pipetting, assembly) needed in lab automation and drug discovery. It introduces Asymmetric Co-Bootstrapping, which uses early behavior cloning to stabilize online RL, then switches to value-based learning to avoid policy drift. ACoB-Stream architecture decouples experience collection from policy updates to handle large VLAs efficiently, achieving 10.9x throughput improvement and 98.3% success on chemistry tasks. This matters directly: robust, fast fine-tuning of VLAs for precise robotic actions could automate wet-lab experiments at Isomorphic Labs, reducing human error and accelerating hypothesis testing in drug discovery pipelines.

Antoine Lorentz, Stéphane May, Valentine Bellet, Dawa Derksen · hf_daily_papers

Diffusion models are now being repurposed as elevation-map refiners: a modified Stable Diffusion 3, conditioned on both noisy photogrammetric DSMs and Pléiades satellite imagery, cuts dense-urban elevation RMSE from 6.00 m to 3.45 m, and generalizes to a held-out city (Bordeaux) with similar gains. The key technical moves — pruning the text stream and applying patch-wise normalization — let a natural-image pretrained model transfer to LiDAR ground truth without catastrophic forgetting. This points toward a cheap alternative to aerial LiDAR: instead of paying for high-accuracy sensors, you can clean noisy stereo-derived 3D maps with a generative prior. For your geospatial background, it’s a strong example of foundation-model priors replacing hand-tuned classical pipelines, and the multimodal conditioning recipe is directly reusable in other remote-sensing tasks.

reddit_ml

A comprehensive, open-source AI engineering curriculum (523 lessons from linear algebra to production serving) is now available as downloadable EPUB/PDF books and in eight languages. The stdlib-first code means you see every implementation step rather than relying on library abstractions—ideal for building deep, debug-friendly understanding of the full stack. With CI-validated lessons, it's a high-quality resource for self-study or teaching, directly applicable to your ML engineering work.

reddit_ml

A community benchmark shows a quantized 8B vision-language model running locally can beat GPT-5.6 Terra on structured tax-form extraction (21/32 vs 7/32 W-2s) while collapsing on Indian bank statements because dd-mm-yyyy gets parsed as mm-dd. The lesson isn't that small models are ready to replace frontier APIs — it's that narrow document-understanding tasks are becoming tractable at the edge, and that failure modes are brutally culture-specific. Also notable: the default qwen3-vl:8b tag in Ollama is the thinking variant, which burns its entire context on long contracts and returns nothing; the instruct variant is required. Self-verification barely changed outputs, and published OCR ground-truth sets contain wrong answers. For ML engineers building production extraction pipelines, this reinforces the value of fine-tuning small models on your own document distribution rather than relying on generic reasoning — and the need to test date/locale handling explicitly.

sebastian_raschka

The key strategic shift: with capable open-weight base models available, the highest-ROI work is now in post-training — specifically distilling reasoning chains into more token-efficient inference. Ember-1 demonstrates this by matching or exceeding larger closed models while using significantly fewer tokens per response. That's a direct challenge to the pretraining-scale arms race, suggesting the next frontier is data quality and targeted fine-tuning rather than raw compute. For anyone building production LLM systems, this means you can now achieve frontier-adjacent performance at a fraction of inference cost by starting from existing weights — and it also signals a commoditization of base models, pushing value capture toward application-level optimization and alignment. The inflection point matters for ML infrastructure choices and model selection strategies in the near term.

Xiaochen Ma, Zimo Meng, Junzhu Liang, Youhe Jiang · hf_daily_papers

RayOrch introduces a programming model and distributed engine that preserves parent-child lineage through ordered expansions and gathers, solving a fundamental bottleneck in foundation-model data preparation. By batching variable-cardinality children across parents while maintaining ordering and failure isolation, it achieves 15× speedup on MinerU scaling from 4 to 64 GPUs and reduces end-to-end time by 13–29% over Ray Data and Daft. This matters because data prep pipelines—for documents, videos, or biomedical records—are increasingly the throughput limit for model training. For anyone building large-scale ML infrastructure, RayOrch offers a principled alternative to flat-record systems, with open-source code available. The lineage-aware design is particularly relevant for regulated domains like drug discovery, where traceability and partial failure handling are critical.

Ayush Jain, Sreeharsha Paruchuri, Ishita Gupta, Fan Zhang · hf_daily_papers

TrackEverything solves a fundamental trade-off in point tracking by representing videos as persistent 3D scene tracks in world coordinates, rather than 2D pixel flows. It introduces voxelization-based de-duplication to avoid redundant surface observations, a decomposed endpoint+trajectory refiner that only processes dynamic points, and 3D WAFT to replace expensive 4D correlation volumes. The result: it’s the first tracker that can track all visible points across videos exceeding 1000 frames on a single 40GB GPU, outperforming dense 3D trackers by >20% APD on short clips while matching sparse long-horizon trackers. This matters for geospatial AI and any domain requiring dense, long-term correspondence — think SLAM in dynamic environments, autonomous vehicle scene understanding, or tracking cellular dynamics in drug discovery microscopy. The de-duplication and 3D scene representation insight is architecturally elegant and practically scalable.

Bovard Doerschuk-Tiberi, Yao Yan, Justin Chiu, Hann Wang · hf_daily_papers

A new open platform for LLM evaluation uses head-to-head competitive games that automatically scale difficulty, avoiding benchmark saturation. It tests strategic planning, adaptation, and robustness under uncertainty across Chess, Poker, and Werewolf. For Nathan, this is directly relevant to evaluating foundation models for drug discovery, where decision-making under uncertainty and exploration are critical. The robust, reproducible infrastructure also speaks to ML engineering interests.

Kristian Schwethelm, Daniel Rueckert, Georgios Kaissis · hf_daily_papers

Looped LMs (shared-layer transformers) can dynamically allocate compute per token — easy tokens exit early, hard ones loop more. The catch: standard batching breaks because tokens diverge on loop count. This paper introduces Continuous Depth Batching, a scheduling system that pools tokens at each loop step, pre-allocates KV cache for looping tokens, and predicts exit points to prepare async batches. On Ouro and Huginn models, CDB achieves up to 99% of the theoretical speedup from depth-adaptivity. The main bottleneck shifts to non-looped layers (embeddings, LM head), suggesting fully looped architectures are the design sweet spot. For you: this directly impacts inference efficiency — a key concern for running large models in production (e.g., at Isomorphic for docking or sequence prediction). If looped architectures become viable, they could dramatically cut per-token cost without sacrificing quality on hard examples, which matters for both your work and any AI startup building on foundation models.

ern_blog

Bill Bengen (of the 4% Rule) now recommends a 55/40/5 portfolio with equities split equally across US large, mid, small, micro-cap, and ex-US stocks, claiming safe withdrawal rates of 4.7–5.5% (and potentially 7% after valuations normalize). ERN's historical simulations pour cold water on this — the higher withdrawal rates come with meaningful failure risk and rely on vague 'flexibility' assumptions. For anyone in the FIRE camp, the takeaway is that the 4% rule's original logic still holds; shifting to a more diversified equity slice doesn't magically unlock higher spending without accepting deeper cuts in bad sequences.

abnormal_returns

Bond yields at 5%+ have shifted the landscape for fixed income investors. The consensus from multiple sources: bonds are now investable again, with clear implications for portfolio construction, especially for those seeking income or diversification. For your personal portfolio, reconsider bond allocations in your ISA/SIPP given the current yield environment — rising yields also introduce equity market risk that you should monitor.

wealth_common_sense

The macro environment is throwing contradictory signals: mortgage rates above 7%, oil at $100, and inflation at 3.4% while unemployment remains at 4% and bond yields offer 5%. This is the kind of stagflationary mix that historically punishes both stocks and bonds, making traditional 60/40 portfolios less effective. For you, as a FIRE-focused index investor, this means revisiting your asset allocation — consider tilting toward inflation-linked bonds, commodities, or cash proxies inside your ISA/SIPP to hedge duration risk. Rising rates also depress growth stock valuations, so your tech-heavy holdings (including your own employer's equity) may face headwinds if this persists.

abnormal_returns

The structural case for index investing continues to strengthen: most stocks underperform and passive strategies have rightly dominated. Meanwhile, the massive AI capex buildout will eventually convert to depreciation, a future headwind that could affect broader market returns. On the personal front, the enduring lesson from Buffett is to align your strategy with your structural advantages rather than chasing stock picks. The broader theme of 'staying private longer' applies both to portfolio construction (low turnover, long holding periods) and corporate strategy (delaying IPOs to avoid short-termism).

wealth_common_sense

The structured note market — long the domain of private banks and wealthy clients — is being repackaged into ETF form, democratizing autocallable income strategies. The key reframe: treating autocallable payoffs not as yield enhancement but as insurance underwriting, where you're systematically selling downside protection on single stocks or indices for premium income. This matters because it changes the risk profile of what retail income investors can access: you're not buying bonds, you're writing contingent options with path-dependent outcomes, and the ETF wrapper adds liquidity and transparency to what was historically an opaque, high-fee product. For a FIRE-focused index investor, this is worth understanding mainly as a cautionary tale — these strategies can look attractive in low-volatility regimes but carry tail-risk asymmetry that most accumulation-phase investors should avoid. But if you're tracking financial product innovation, this is a notable shift in how income is manufactured and distributed.

hacker_news

Hacker News is ablaze with a post questioning Google's recent bizarre behavior, from shoddy AI summaries to unwieldy search results. The community's deep dissection unearths a growing unease: Google's once-vaunted engineering culture may be succumbing to bureaucratic bloat and a misguided rush to ship AI features. For you, as an engineer inside Alphabet, this is more than noise—it's a reflection of internal tensions that could directly impact your team's autonomy, the quality of your tools, and the company's external reputation. The broader lesson: when a tech giant prioritizes product velocity over coherent user experience, it erodes trust and opens the door for smaller, nimbler competitors.

sifted

Cohere's chief AI officer warned that without local investments, Europe will be strategically dependent on US and China for AI — held 'by the throat'. This amplifies calls for a coordinated European AI strategy, especially around compute and talent. For Nathan, working in London at a cutting-edge AI drug discovery firm, this trend affects the local talent pool, startup ecosystem, and potential for public-private partnerships. It also signals that AI geopolitical tensions are heating up, with implications for international collaborations and data sovereignty. Notably, it reinforces the importance of strong homegrown AI capabilities, which could benefit companies like Isomorphic Labs as they recruit and collaborate in the region.

hacker_news

A dispute over a billion dollars in Nvidia stock exposes the extreme financial leverage and risk embedded in AI hardware equity. This case reinforces how startup viability can hinge on Nvidia's market position and stock compensation structures, a critical consideration for any AI-native company. For Nathan, it underscores the strategic importance of aligning supply chain and equity arrangements with the realities of Nvidia's dominant role in the AI ecosystem, directly relevant to capital-intensive ventures like drug discovery.

tech_eu

European tech funding remains robust with over €2B across 65+ deals last week; robotics, AI, and healthtech led, and Portugal topped country rankings thanks to TEKEVER's $580M Series D. For you, the standout is Basecamp Research's $140M Series C backed by Anthropic and Nvidia — a direct nod to AI-driven genomics for drug discovery, reinforcing that AI labs are placing strategic bets on biotech data infrastructure. Also notable: Biolevate raised €30M for life sciences AI, and Bird's debt restructuring (headcount 1,000→120) signals a broader shift toward capital-efficient AI-first plays. The healthtech+AI investment wave aligns with the competitive landscape you watch.

hacker_news

Dario Amodei took AI safety arguments to a mainstream comedy audience, signaling that the Overton window for discussing existential AI risk has shifted. For Nathan, this matters because public perception and regulatory pressure directly affect the operating environment at Isomorphic Labs — even if the R&D focus differs from general-purpose AI. The fact that a frontier lab CEO chose SNL over a technical conference suggests alignment and safety messaging is now a strategic priority, not just an engineering concern.

hacker_news

Fireworks AI released Ember-1, an open-weight 8x7B MoE model that outperforms GPT-3.5 and Mixtral on key benchmarks while being significantly cheaper to serve. The key insight isn't just the performance—it's that a relatively small team achieved this by focusing on data quality and training curriculum rather than compute scale. For your work at Isomorphic Labs, this reinforces that model efficiency breakthroughs are happening faster than expected, which directly impacts how you think about serving custom models for drug discovery. The sub-7B active parameter count means this could run inference on modest hardware, potentially enabling on-premise or edge deployment of specialized biological LLMs without massive cloud GPU costs.

huggingface_blog

Holo4 represents a shift toward generalist computer-use agents, moving beyond narrow task-specific automation. By training on diverse GUI interactions, it enables models to understand and operate software across different environments. This matters for scaling agentic AI in real-world workflows — whether that's automating data pipelines, managing cloud consoles, or even orchestrating complex lab information systems. For Nathan, the underlying techniques (likely leveraging vision-language models and reinforcement learning) could inform how we build agents that interface with scientific software and internal tools at Isomorphic. The jump from brittle, scripted automation to adaptable generalist agents is a key step for practical AI deployment.

cloudflare_blog

Cloudflare shipped an experimental Emscripten target for wasm-bindgen, letting native Rust code and even Tokio-based async apps run unmodified on Workers. This effectively virtualizes platform features like timers, filesystem, and TCP sockets on top of Workers' Node.js-compatible APIs. The proof point: they got a Rust-native Minecraft server (Pumpkin) running inside a Durable Object with real TCP ingress. For platform engineers, this collapses the gap between edge and traditional compute far more than raw WASI ever did — you can now run dependency-heavy Rust crates, async runtimes, and real network sockets at the edge without rewriting for Worker constraints. Worth watching if you're building latency-sensitive ML inference or geospatial pipelines that need to move close to data. Still pre-release, but the architectural direction is clear.

cloudflare_blog

Cloudflare's Kitesurf — an agent-native browser built on Workers — is maturing beyond a gimmick. The key shift is WebMCP support: sites can now expose typed functions (searchFlights(), set-location()) directly to agents, bypassing fragile pixel-clicking and DOM scraping entirely. Combined with new CSS layout and CSSOM API coverage, Kitesurf is becoming a viable runtime for agentic tasks that require real rendering fidelity, not just headless DOM access. The strategic tell: Cloudflare is betting that the browser becomes an execution environment for AI agents, with the website owner opting in to expose capabilities. For an ML engineer watching agentic infrastructure, this is a meaningful data point — the bottleneck isn't just model reasoning, but reliable grounding in live web environments. If agents can call explicit APIs instead of reverse-engineering UIs, downstream reliability and cost improve dramatically. Worth a quick look if you're tracking agent platforms or building on browser-APIs.

cloudflare_blog

Cloudflare open-sourced Forge, a pluggable code-generation pipeline that turns OpenAPI specs into SDKs, CLIs, docs, MCP servers, and even Cap'n Web RPC bindings — designed to run in CI with per-change preview builds across hundreds of repos. The key insight is treating agents as first-class customers: every product now needs machine-readable surfaces, not just human docs. For Nathan, this matters as an infrastructure pattern for large ML platforms — if Cloudflare can generate consistent client surfaces across 3,500+ operations and multiple languages with distributed ownership, similar pipelines could standardize model-serving APIs or feature-store contracts at Isomorphic. It also signals a shift in how API teams can reduce cross-team coordination overhead, a problem Nathan likely recognizes from Lyft's geo/platform work. Early, but worth watching for its pluggable transformer model and preview-first workflow.

cloudflare_blog

Cloudflare's 2026 letter reveals two inflection points: AI-aided 'vibe coding' has triggered a web renaissance, with 7M+ developers now building on their platform, and automated traffic surpassed human traffic in May 2026—five years ahead of prior forecasts. The real insight is the emerging tragedy of the commons: as AI agents consume vast content without paying, they risk concentrating visibility on established players, making it harder for new creators to be discovered. For platform engineers and startup watchers, this signals a shift where distribution cost becomes the critical bottleneck, not creation.

stat_news

Anthropic identified five instances where its models were used to support work that could aid biological weapons development — including gain-of-function research proposals and experiments to adapt bird flu for mammal infection. The company blocked these requests but couldn't determine intent. This illustrates the fundamental dual-use dilemma: the same AI capabilities that accelerate drug design can also be weaponized. For someone working at the intersection of AI and drug discovery, this matters because it signals that forthcoming regulation around AI in biology will likely tighten — and that the burden of demonstrating safe intent will fall on legitimate researchers. The article reinforces why alignment and guardrail research is not just an AI safety abstraction but a concrete operational risk for any AI-native pharma company.

biopharma_dive

Merck and Daiichi have scrapped plans for accelerated FDA approval of their lung cancer ADC, I-DXd, after regulators signaled the current data isn't strong enough. This eliminates a strategic advantage in the fiercely competitive ADC landscape, where rivals like AstraZeneca and others are advancing rapidly. For Nathan, it's a reminder that regulatory risk remains high even for promising candidates, and that robust clinical evidence is non-negotiable. The setback also highlights the importance of differentiation in ADC design—an area where AI-driven approaches could eventually tilt the playing field.

biopharma_dive

The next competitive moat in biopharma won't be just novel targets or AI models—it will be the talent and systems that efficiently translate discovery into scalable, manufactured therapies. This reinforces that Isomorphic Labs' long-term value depends as much on building robust ML infrastructure and platform engineering as on the underlying drug discovery science.

stat_news

Trump is leveraging contested executive authority to impose additional funding cuts on HHS, likely targeting NIH grants, FDA operations, and public health programs. This accelerates a trend of reduced federal support for biomedical research and regulatory capacity, which could slow drug approval timelines and shrink the pipeline for early-stage biotech — directly impacting the funding environment and strategic landscape for AI-driven drug discovery companies like Isomorphic Labs.

stat_news

GLP-1 marketing is pivoting from obesity-as-disease to lifestyle enhancement, as Novo Nordisk and Eli Lilly face stalled insurance coverage and sell directly to consumers. This shift reflects the commercial reality that payer adoption has plateaued due to high prices and broad usage. Separately, Merck and Daiichi Sankyo withdrew their U.S. accelerated approval application for an experimental small-cell lung cancer therapy after the FDA deemed mid-stage data insufficient — the second setback in their cancer alliance. The pullback underscores the high bar for ADC-based therapies in aggressive indications. For you, these stories highlight how commercial strategy (GLP-1) and clinical risk (ADC setbacks) shape the pharma landscape, relevant for understanding industry dynamics and competitor portfolios.

stat_news

Novo Nordisk is repositioning GLP-1s like Wegovy from clinical obesity treatments toward lifestyle wellness products, dropping “weight management” language in favor of “weight loss” and referring to users as “customers” rather than “patients.” This shift is driven by the need to expand beyond insured chronic-disease populations into a broader, cash-pay wellness market. The implications go beyond marketing: it changes the regulatory narrative, pricing strategy, and competitive dynamics for next-generation metabolic drugs. For AI-driven drug discovery, this commercial tailwind makes obesity and metabolic targets even more attractive, but it also raises questions about long-term payer pushback and the valuation of lifestyle-driven indications versus medical ones.

biopharma_dive

The bar for drug success is moving beyond safety and efficacy into the lived experience of patients. Companies are now redesigning therapies to minimize dosing frequency and shrink time-to-administration, treating convenience as a core product attribute rather than an afterthought. This reflects a market where adherence and real-world usability directly impact commercial viability and reimbursement. For AI-driven drug discovery, this means optimization functions should extend to patient-centric parameters — like dosing schedules, formulation stability, and delivery ease — not just biochemical binding. Predictive models that can simulate real-world adherence patterns will become a competitive edge, especially in chronic and pediatric indications where the burden is highest.