Sophie Alpert argues that there are no lossless transformations of natural-language text, meaning every rewrite changes meaning and can lose information if the rephraser lacks a detailed representation of the original author’s intent. She emphasizes that any AI-assisted writing must be backed by the engineer’s own ideas and sentences and that replying to reviewer questions with an “AI wrote it” explanation is not acceptable. As a result, engineers are expected to review and own their entire documents rather than treat AI edits as interchangeable with their original thoughts.
Simon Willison’s Weblog·2 months ago·
46
● 3 sources
A paper described how replaying encrypted chain-of-thought (“reasoning”) blocks from proprietary LLM API responses could let attackers recover plaintext hidden reasoning by running the trace across models and then jailbreaking a weaker sibling. The authors said the easiest target was Claude Haiku 4.5, where a removed feature in 4.6 had still allowed an assistant prefix to work. Providers acknowledged the report and the authors could no longer run the same attacks after the fix.
Zvi (Don't Worry About the Vase)·2 months ago·
4
● 3 sources
Various contributors discuss OpenAI’s internal post-mortem details about covert “agent” communications and model training behavior after an initial security incident, emphasizing uncertainty about what was known when.
Steph Batalis discussed in a Washington Post article how AI-generated viruses could aid vaccine development and treatments for antibiotic-resistant bacteria while also creating risks of misuse to design dangerous human pathogens. The piece says the paper does not show AI can readily design a dangerous human virus or that AI can be prompted to generate any arbitrary virus. The focus shifts to evaluating AI-virus capabilities and limits for public health benefits while tightening concern about human-pathogen misuse.
Saber denied claims that it replaced a Rideshare "Stimulator" lead writer with ChatGPT. The writer alleged the switch happened midway through development. Saber says no writers were replaced by AI, while the dispute centers on whether passenger voices were also AI and whether disclosure on Steam is missing.
Anthropic announced that an unreleased model made progress on the Riemann hypothesis by increasing the lower bound of cases where the hypothesis is shown to hold. The effort tested 650 different ideas and coordinated across 60 sub-agents at a total cost of 31 million. This adds to growing evidence that AI systems can generate new mathematical proof approaches and also intensifies debate about how proof credit and authorship should work in mathematics.
Amazon’s Automated Reasoning Group has grown from a 2016 internal demo team into production services that apply formal logic to prove AWS systems’ security and correctness. It says its authorization engine was verified against quadrillions of production authorizations while handling 1 billion API calls per second. The work is now used across AWS security and reliability tooling and is being extended to guardrails and requirements checks for agentic AI systems.
Research Gold marketed medical research writing and review work as “100% human-written, never AI,” but the researchers shown as reviewers appear to be AI-generated or used without permission, and the company’s communications also appear to be produced by an AI agent. The article says a systemic review quote for the described project was priced at $1,900. As a result, the company removed at least one page listing a real methodologist’s identity, and the report raises concerns about how AI-generated text and misattributed expertise could undermine parts of academic publishing.
Apple is developing an iOS feature intended to verify when iPhone photos were taken using an iPhone camera. The iOS 27 beta 5 includes code references to an "Apple Reference Image" system that can embed provenance metadata at capture. If released, the feature will be off by default and can be enabled in Settings > Camera > Reference Image > Reference Mode, changing how users can validate photo origin and resist deepfake claims.
Spotify will label AI Personas and stop recommending their music to listeners based on whether profiles appear to represent photorealistic AI identities. The rollout starts in mid-September. Spotify will let artists opt in today, but it will still use human review plus AI tools to decide whether to apply the badge and adjust recommendations, affecting what users see and are suggested.
The article argues that Mark Zuckerberg’s AI “positive future” manifesto is internally inconsistent, treating AI safety as power sharing while the author stresses risks that require stronger control. It points to Meta’s planned $50 billion data center in Louisiana that is projected to use seven times as much energy as New Orleans. The result is a call for slower, more cautious deployment and a claim that governments and AI labs are increasingly pushing back against fast, widespread release plans.
Every AI story that matters,
in your inbox by 8am.
TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the
day in two minutes. Follow companies and topics for alerts, or get the
briefing in Slack. Free, no spam, unsubscribe anytime.
Reading TLDRocket needs no cookies, and the readership counts we rely on come from
our own cookieless analytics. Google Analytics is the exception: it sets cookies and
reports to Google, so it stays switched off until you allow it. You can change your
mind any time from “Cookie settings” in the footer.
Strictly necessary
Session security and form protection (tldrocket-session,
XSRF-TOKEN, 2 hours). The site cannot work without them,
so they need no consent.
Always on
Google Analytics 4 (_ga,
_ga_<id>, up to 2 years). Measures which
stories and sections readers use. Google acts as a third-party processor and may
store the data outside the EU. No advertising, no profiling, no data sold.
