OpenAI to Governor Abbott: we’ll pay for our own power and water
OpenAI published a letter to Texas Governor Greg Abbott on August 10, committing to pay for its own infrastructure, support new electricity generation in Texas, and minimize water use as it builds out data centers in the state. The backdrop: Abbott has set conditions for data centers, including paying for their own grid connections, reusing water where possible, not driving up residential electricity prices, and forgoing taxpayer-funded incentives; Meta made a similar commitment the same day (see local coverage). The telling part, as I read it, is the addressee: with federal AI legislation stalled, the constraints that actually bind projects like Stargate are state-level power and water, so companies are settling that account with governors.
A think tank turns the pacing letter into 23 recommendations, with budget numbers
The Institute for Progress published “How Should the US Prepare for Increasingly Automated AI R&D?” on August 6, a direct response to July’s Pacing the Frontier open letter signed by more than 1,300 frontier lab employees. I wrote a breakdown of that letter (in Chinese); my main open question was what the “ability to pace” would actually consist of. The report offers 23 recommendations in seven categories, and the distribution is itself a statement: seven on extending US AI advantages, six on verification technology, competition and verification side by side. The specifics come with money and institutions attached: at least $84 million a year for CAISI (the Commerce Department’s Center for AI Standards and Innovation, the renamed US AI Safety Institute), an industry consortium to prototype technology for verifying how compute is actually used, and screening requirements for DNA synthesis providers. This reads like the engineering appendix the letter was missing: verifiable compute monitoring first, pacing second.
Stanford HAI: before regulating mental health AI, define it
Stanford HAI researchers, writing up a June policy workshop, lay out three gaps in governing mental health AI. The definitional one is the sharpest: general-purpose chatbots, companion bots playing therapist, wellness apps, and clinical tools are different things, and a broad ban on “AI therapy” can push users toward unregulated general-purpose chatbots, the opposite of the intent. Evaluation lags too: the highest-stakes conversations (such as self-harm crises) are rare and hard to simulate, and the real chat data needed to study safety at scale sits inside companies, with no structures for sharing it with outside researchers or regulators. Nor can company staff simply read it themselves: standard privacy practice at large platforms keeps employees from viewing user conversations directly, so what’s workable internally is mostly aggregate-level features. Current metrics reflect developer priorities rather than therapeutic outcomes. By the article’s count, more than 140 mental health AI bills have been introduced at the state level, while federal work stays narrowly focused on minors. The authors’ judgment: disclosure of AI identity, crisis protocols, and data protection have consensus and can move now, but the conflict between engagement-optimized business models and user well-being is one that disclosure requirements won’t fix.
SIGIR paper: RAG poisoning makes a model’s attention collapse onto the poisoned documents
RAG poisoning inserts crafted malicious documents into the corpus a model retrieves from, steering its answers. This SIGIR 2026 paper finds that output-side detection is unreliable: poisoned answers actually score lower perplexity (they read as more fluent) than benign ones. The internal signal is stronger. When a generation is being manipulated, attention concentrates abnormally on the few injected documents and attention entropy drops; the authors name this document-level attention collapse and build a detector, D-SCAN, on top of it. Detection moves from checking the answer to watching the model’s internals, and it catches attempts even when the attack fails to change the final answer, which matters if you want to monitor poisoning attempts in production rather than audit answers after the fact.
Fake news video is entering the pure-synthesis era, and the old detection signal is gone
Fake news video research to date has centered on cheap fakes: real footage recycled under a misleading caption, where detectors catch the mismatch between what the video shows and what the text claims. This ACM Multimedia 2026 paper points out that text-to-video models generate the footage from the caption, so image and text match by construction and the mismatch signal has nothing to catch. The authors build PS-FNVD, which they describe as the first pure-synthesis fake news video dataset, and recast detection as three-way classification (real / cheap fake / pure synthesis). Their framework combines two kinds of signals: semantic reasoning about whether the content itself holds up, and the physical traces generative models leave in the footage, meaning visual giveaways such as lighting and surface-texture anomalies, discontinuities between frames, distorted bodies or faces, and motion that breaks physics; a multimodal model checks sampled frames along these dimensions. Accuracy beats the runner-up baseline by about 12 points. For content moderation this is a real shift: fully synthetic footage has no earlier source for provenance lookup (where did this footage first appear?) to trace back to, so detection has to lean harder on the generation traces themselves.
Needle 2: a 14MB on-device agent model for sub-$200 hardware
Cactus Compute released Needle 2, an open 45M-parameter model compressed to a 14MB binary that runs a full session in 28MB of RAM. “Agent model” here is narrow by design: no open-ended chat, just mapping user requests to typed device function calls (tool calling, device control, structured extraction), on hardware like the Raspberry Pi 5 (500+ tokens per second decode, per the vendor), budget phones, and ESP32-class microcontrollers. On the vendor’s Mobile Actions benchmark it scores 63.7%, below the 5x-larger LFM2.5 230M at 69.1%, at 5–70x smaller size; Pebble already ships it in the Index Ring for offline voice-to-action. This is what edge agents look like in practice: rather than squeezing a large model into a phone, train a small model for exactly one job, understanding a command and calling the right function.
Dan Luu: token efficiency is the wrong way to pick a language for coding agents
The popular claim is that dynamic languages like Python cost fewer tokens and therefore suit coding agents better. Dan Luu ran two sizable evals (implementing a zstd decoder from the spec; modifying pandoc against a holdout test set) across a dozen languages including Rust, Go, C++, Haskell, Python, Ruby, and Clojure, measuring both correctness and cost. The advantages seen on trivial tasks did not carry over: at medium reasoning effort dynamic languages held a modest edge, at high effort the results were mixed with static languages pulling even, and outcomes correlated (weakly to moderately) with language popularity more than with type discipline. His practical conclusion: pick a mainstream language, and don’t switch languages over token-count arguments.
Python’s main crypto library has post-quantum algorithms now (and has since May)
This one is not from this week: Trail of Bits shipped NIST’s post-quantum algorithms ML-KEM (key encapsulation) and ML-DSA (signatures) in pyca/cryptography version 48, released in early May and written up by the team on June 30, and it circulated again this week via Bruce Schneier. pyca/cryptography is the 11th most-downloaded package on PyPI at 1.2 billion downloads a month, and much of the Python ecosystem’s encryption, signing, and TLS ultimately rests on it, so the whole dependency chain above it can now start migrating. The costs are plain: one ML-DSA-65 signature is 3,309 bytes against Ed25519’s 64, so protocols and storage formats need room for signatures roughly 50x larger. The deadline pressure comes from harvest-now-decrypt-later: encrypted traffic recorded today can be decrypted once quantum computers mature, so the switch has to happen well before then; the White House has ordered high-value and high-impact federal systems to complete the transition by the end of 2030 for key establishment and the end of 2031 for digital signatures.
Today in one line: AI governance is trading principles for line items: “pacing” became 23 recommendations with budget numbers, and “responsible infrastructure” became a letter promising to pay the power and water bills.