Thursday, 25 June 2026

OpenAI ships first custom inference chip; Anthropic accuses Alibaba of model distillation via API; Microsoft quantum breakthrough disputed on Python errors

Today's Lead

Engineering

TechCrunch

OpenAI Unveils Its First Custom Chip, Built by Broadcom

OpenAI unveiled Jalapeño, its first custom inference processor co-designed with Broadcom, optimized for inference workloads with substantially better performance-per-watt than general-purpose GPUs. The strategic logic is identical to playbooks pioneered by Google (TPUs) and Amazon (Trainium/Inferentia): at sufficient inference scale, the margin math forces vertical integration. OpenAI has been unprofitable despite enormous revenue in part because GPU inference costs dominate; a custom chip that cuts inference cost per token is a structural fix no amount of pricing or operational efficiency can match. For engineers building on OpenAI's APIs, the short-term effect is neutral — Jalapeño remains in testing. The long-term implication is that OpenAI can optimize inference economics in ways opaque to competitors and make latency-versus-cost tradeoffs unavailable on commodity hardware, producing a structural advantage if the chip matures. The timing also signals OpenAI's confidence that its inference volume is large and durable enough to amortize hardware NRE — which is itself a claim about the business.

Read →

Engineering

Reuters

Anthropic Says Alibaba Illicitly Extracted Claude AI Model Capabilities

Anthropic has accused Alibaba of using Claude API outputs to train competing models through model distillation — accessing Claude at scale through Chinese resellers offering discounted tokens at 70-90% below official pricing, then using those outputs as training data for student models. The technical operation is straightforward: call the teacher model at volume, use its outputs to supervise a cheaper model, and the student inherits a substantial share of the teacher's capabilities at a fraction of the training cost. The legal status is genuinely contested — distillation from another company's API while violating terms of service is different from distillation of publicly available outputs, but the IP theory that API outputs are protectable trade secrets hasn't been tested at scale. The irony raised by commenters is real: Anthropic trained Claude on copyrighted text without explicit authorization, making the line between legitimate and illegitimate capability transfer difficult to draw with philosophical consistency. Whatever the legal outcome, the episode demonstrates that API access at scale to frontier models is a structural vector for capability transfer, and that pricing alone — without rate limits, usage monitoring, and reseller controls — cannot prevent it.

Read →

The Register

Boffin Claims Microsoft's 'Quantum Leap' Is Invalid Due to 'Basic Python Errors'

Dr. Henry Legg at the University of St Andrews published a peer-reviewed critique arguing that Microsoft's February 2025 announcement of a Majorana topological quantum computing breakthrough is invalidated by code-level errors in the Python analysis: improper array filtering and incorrect voltage antisymmetrization selectively surfaced favorable results while concealing alternatives that would have failed Microsoft's own evaluation protocol. If sustained, this critique resets quantum computing timeline expectations considerably — Microsoft's announcement had generated substantial investor and policy attention around near-term practical quantum capability. The deeper issue is not the Python errors themselves but what they reveal about research incentives: when a result has enormous strategic and financial value, the analysis pipeline rarely receives the same scrutiny as the experimental apparatus, and code correctness is almost never audited by peer review. For engineers evaluating quantum computing timelines for planning purposes, the most defensible heuristic remains unchanged: treat any specific breakthrough claim as provisional until independently replicated with the analysis code open for inspection. Microsoft has not yet formally responded.

Read →

Armin Ronacher

The Coming Loop

Armin Ronacher (creator of Flask and Jinja) argues that AI-driven autonomous coding loops — systems that continuously iterate on code without human intervention — are becoming technically inevitable under competitive pressure, regardless of engineers' legitimate concerns about comprehensibility and maintainability. His key empirical observation is that these loops excel at bounded, verifiable tasks (porting, optimization with measurable benchmarks) but generate defensive, over-complex code with poor invariants when applied to open-ended systems problems — producing codebases that become opaque enough that only AI can diagnose them. This creates a self-reinforcing dependency: the more AI-generated code degrades human-readable invariants, the more work requires AI assistance to navigate, which increases pressure toward autonomous loops. Ronacher's argument is not that loops are good but that engineers who refuse to understand them are ceding the design conversation to those who don't share their concerns. For engineering leaders, the implication is that the strategic question is not whether to engage with autonomous coding but how to bound it — preserving human-intelligible system-level invariants even when component code is AI-generated, and determining which tasks are genuinely bounded enough to be safe candidates.

Read →

LWN.net

Free-Threaded Python: Past, Present, and Future

Python 3.13 introduced experimental GIL removal (free-threading), enabling true parallel thread execution; performance overhead has improved from 20-40% in 3.13 to 0-10% in 3.14 through a unified ABI design allowing extension modules to support both GIL and free-threaded builds from a single codebase. The core engineering challenge reflects Python's architectural roots: nearly every object is shared mutable state, requiring memory management solutions like biased reference counting and quiescent-state-based reclamation that avoid per-operation locking overhead. CPython developers project free-threading as the default between Python 3.16 (2027) and 3.20 (2031), with the GIL-based interpreter phasing out after 2030 — over 50% of top PyPI packages are already preparing for the transition. The practical consequence for engineers building CPU-bound Python services: free-threading removes the fundamental reason that serious parallel workloads historically required multiprocessing (high memory overhead) or abandoning Python for data-parallel sections. The remaining risk is ecosystem: C extensions that relied on GIL-enforced atomicity without explicit locks will silently become incorrect, and auditing for this is non-trivial in large dependency trees.

Read →

Crooked Timber

AI Electricity Use: A Lot or a Little?

The piece disentangles two questions that are routinely conflated in AI energy discourse: whether current AI electricity consumption is a problem (probably not — roughly 1% of US demand, broadly manageable), and whether AI company valuations are grounded in achievable revenue (probably not). To justify OpenAI's current valuation on conventional revenue multiples would require revenues 50× higher than present — which, at plausible usage economics, would demand electricity equivalent to roughly half of total US consumption, a physical impossibility on any near-term grid. Market saturation is the second constraint: OpenAI already has 900 million weekly active users, and price elasticity makes the arithmetic of 50× growth implausible regardless of energy supply. The analysis also notes complications in the 'modest impact' framing: agentic AI is significantly more token-intensive than chat interfaces, training costs are ongoing rather than one-time, and infrastructure burden falls disproportionately on regions and nations with inadequate grid capacity. The policy implication is that AI energy debates often conflate a legitimate but manageable infrastructure question (can we power current AI loads?) with a financial question (are valuations justified by achievable demand?) — and the second is the one that should be driving more concern.

Read →

Marginal Revolution

What Should the UAP Scientific Advisory Board Do?

Tyler Cowen argues that the UAP Scientific Advisory Board's principal failure mode is scope expansion — science committees reliably produce credible outputs when constrained to concrete, data-driven verification questions (do reported videos have corroborating radar, satellite, and independent eyewitness evidence?) but tend toward dysfunction when asked interpretive questions before first-order factual questions are established. The institution-design prescription is to constrain scope explicitly: build credibility on what is directly measurable and verifiable, establish the factual baseline, and explicitly defer interpretive questions to a later stage. This is a specific instance of a general pattern in advisory science: when committees are given questions they cannot yet answer with available data, they produce outputs that mix fact with speculation in ways that undermine both the credibility of the committee and the credibility of the underlying evidence. For engineers familiar with systems decomposition, the argument maps onto the problem of mixing requirements specification with implementation decisions before the problem statement is stable — the fix is structural, not a matter of asking the committee to reason better.

Read →

Marginal Revolution

CoPaper.AI and the Automation of Empirical Research

Tyler Cowen examines CoPaper.AI, a Stanford-associated project claiming to automate empirical research paper generation from raw datasets in under 30 minutes — converting data directly into a complete paper with code, visualizations, and written analysis. The claim, if it holds at quality thresholds that satisfy peer review, would compress a substantial portion of routine empirical social science production from weeks to minutes, raising immediate questions about what judgment currently embedded in that process is genuinely valuable versus merely expensive. The automation-of-research question is structurally parallel to the automation-of-coding question: if the bottleneck shifts from data analysis to question formulation and result interpretation, what does that imply for how graduate training, peer review, and scientific credit allocation should be restructured? Cowen notes the capability is currently overstated in its marketing but directionally real. For researchers and research managers, the useful frame is not 'will AI automate research?' but 'which parts of the research production process are genuinely requiring expert judgment versus executing a well-specified pipeline?' — the answer to that question shapes where human effort should concentrate.

Read →

Humanities

JSTOR Daily

Rollerena: New York's Fairy Godmother

Rollerena was a queer performance artist who roller-skated through New York starting in 1972, dressed in elaborate gowns and rhinestoned glasses, becoming a celebrated presence in queer nightlife before evolving into a political force during the AIDS crisis — joining ACT-UP and earning the designation 'Fairy Godmother of the AIDS activist movement.' The JSTOR Daily piece argues that creative visibility and political activism were not parallel projects but the same project: Rollerena's conspicuous presence in public space was itself a political act during decades when queer existence was systematically criminalized and targeted. The historical arc captures something specific about how queer counterpublics formed before networked communication: through embodied presence in contested physical spaces, building solidarity through repeated, visible coexistence rather than online organizing. The pattern generalizes: performance as political assertion, and the transformation of an aesthetic gesture into a survival strategy when communities face existential threat, is a recurring structure in the history of marginalized groups. Rollerena died in 2021; the piece is also a record of how quickly the people who built the social substrate of a community can pass out of institutional memory.

Read →

Also today

Intercept Fund Blog

Ending Respiratory Infections

The Intercept Fund, a new $500M initiative, proposes an elimination strategy for respiratory infections combining two technology classes: broad-spectrum preventatives (BSPs) — vaccines, nasal sprays, and antivirals targeting multiple respiratory viruses simultaneously — and air cleaning technologies (ACTs) including filtration and far-UVC light deployed in shared indoor spaces. Modeling shows that neither approach alone achieves elimination due to adoption ceilings; combined at realistic uptake rates (60% BSP, 50% ACT), transmission can fall below the reproductive threshold needed for disease elimination. The strategy is newly feasible because COVID-era platform technologies — mRNA, protein design tools — make rapid BSP development practical in ways unavailable before 2020. The economic case is straightforward: the annual burden of respiratory infections is approximately $600 billion globally, historically underfunded relative to its share of disease burden. For engineers familiar with defense-in-depth reasoning, the combinatorial logic is familiar: two imperfect defenses with independent failure modes can achieve reliability neither achieves alone. The behavioral constraint (adoption is voluntary, not mechanical) is the genuine uncertainty, not the underlying technology.

Read →