Gemini 3.8, Muse Spark 1.3 🤖, Copilot cut costs by removing dead formatting ✂️, workslop is a DoS attack 🌊

Google put out two Flash models on 2 September, its third Flash release in six weeks: Gemini 3.8 Flash at the same introductory price as 3.7, and Gemini 3.8 Flash Cyber, restricted to trusted defenders through a new Fairwind Program. Chrome's security team measured 2.6 times more correct patches than much larger commercial models. On the external CWE-Bench, though, the Cyber model scores 47.2 per cent against a leading frontier model's 47.8, at lower cost.

Meta rolled out Muse Spark 1.3 today in Muse Code and the Meta Model API. In comparisons run by Meta's own engineers, the model needed roughly 20 per cent fewer tool calls and 25 per cent fewer tokens than Muse Spark 1.2 to finish the same coding tasks. It also asks clarifying questions on ambiguous prompts, confirms before consequential actions, and now follows multi-step instructions without dropping constraints.

OpenAI's upcoming Astra model will reportedly use a looping reasoning technique, recurrent depth or opaque recurrence, instead of a visible step-by-step chain, The Information reported Tuesday. Redwood Research chief executive Buck Shlegeris said he is extremely concerned, warning heavier use could totally destroy chain-of-thought monitoring. Anthropic and Google DeepMind are reportedly already discussing the same approach, and Redwood's Ryan Greenblatt warned it could scale toward models reasoning almost entirely outside human inspection.

The Justice Department told a Manhattan federal court on Tuesday that judges should find OpenAI did not breach copyright law using New York Times and other publishers' articles to train its AI, its first intervention in an AI copyright case. It argued training transforms copyrighted work into new material permitted under copyright law, and that AI's benefits far outweigh any competitive harm. The Times said the government was siding with trillion-dollar AI companies over American creators.

AI coding startup Cognition AI is set to close a new funding round valuing it at around $47 billion, according to people familiar with the matter. It is raising around $1 billion, though the final size may grow as it fields what one person called outsize demand, having already drawn nearly $10 billion in investor interest. Terms are still being finalised, the people said.

GitHub Copilot's CLI harness stopped prefixing every file it reads with line numbers, formatting left over from older editing tools that no longer need it. Removing it lowered model-inference cost by roughly 5 per cent in offline benchmarks and cut average daily cost per user by about 3 per cent in a live experiment, with no drop in success rates. The Copilot app and Copilot code review share the harness, so they gain too.

Cline shipped its rebuilt SDK-based harness to 11 million VS Code users by bundling old and new extensions behind a feature flag, working around a Marketplace with no gradual-rollout option. The agent's mistake-limit rate, how often it stops after three consecutive errors, fell from 6.34 per cent of tasks to 0.62 per cent. Reductions ran six to eleven times by model, and Cline calls its own multiples conservative.

XBOW's autonomous security-testing platform kept flagging the same vulnerability multiple times across different endpoints and parameters, burying human reviewers in redundant reports. The fix converts each finding's description, reproduction steps, location and impact into separate numeric fingerprints, weights them differently per vulnerability type, and tunes those weights against findings its own analysts had manually grouped as duplicates. That approach cut reported finding volume by 30 per cent and eliminated 90 per cent of duplicate findings.

An independent researcher built a GPU attention kernel, the code that runs a transformer model's core computation, from scratch in CUDA for Nvidia's B200 chip. Through 14 sequential low-level optimisations to GPU memory layout, loading and compute scheduling, the kernel's speed climbed from a 14.2 per cent baseline to 94.4 per cent of FlashAttention-4's benchmarked speed. The finished kernel was then plugged into the LTX-2.3 video-generation model as a capstone demo.

A study tested Perplexity's sonar and sonar-pro models across 380 software categories and found 59.8 per cent of citations point to domains ranked worse than #100,000. Three affiliated sites, none older than December 2023, have published 215,128 machine-generated best-software pages that the two models retrieved as grounding sources. In two categories, one of the models named a company homepage that actually redirects to an Indonesian gambling portal or a Monaco casino group.

Thoughtworks CTO Rachel Laycock's rebuttal to a DX researcher argues AI exposed, rather than caused, an already broken code review process. The researcher's own numbers show Meta's lines of code per diff reportedly up 106 per cent in a year, and median pull request size up 64 per cent. Her fix is to shift feedback into pairing and design sessions, keeping human review only for sensitive or high-blast-radius changes.

A software engineer argues AI coding agents have killed the human habit of stopping to refactor tangled code. Agents never get lost tracing branches, so they keep adding exceptions to code nobody on the team can still follow, and reviewers rubber-stamp changes they cannot judge. The risk is teams trusting agents precisely because they no longer understand their own code, a mess that also costs agents more tokens per change.

Bruce Schneier published two unsolicited emails from what says it is an autonomous Claude instance, given a server with root, $4.75 of crypto and 24 hours to turn it into $10. In its own account, identity verification blocked it zero times, while captchas, datacentre IP bans and two-day payment settlement stopped it dead. Schneier calls the emails vaguely coherent and draws no conclusion, so the findings stand as an unverified self-report.

A software engineer defines workslop as colleagues pasting AI-generated text at you, an attack that costs almost nothing to produce but real effort to read. His countermeasures include telling junior offenders to stop, treating repeat ones like a slow coding-agent interface, or feeding their text back into an LLM for a summary. For low-stakes cases, like status updates or outside pull requests, he says the simplest fix is to skim, delay reading, or ignore it entirely.

Bryan Roth, a co-inventor of chemogenetics, told a BRAIN Initiative conference on 13 August that he had found seven ongoing Chinese trials of the technique, covering epilepsy, Parkinson's disease and pain. Not all are recruiting yet, and together they add up to roughly ten patients. The method delivers a designer receptor by virus, then switches it on with ultra-low clozapine doses, 1 per cent of the schizophrenia dose in the Parkinson's trials, and still requires brain surgery.

cn is a new Tailwind CSS class-merging engine, built by aidenybai and shadcn as a drop-in replacement for clsx and tailwind-merge with identical APIs. Replaying 144,265 calls harvested from 58 open source codebases, the maintainers measured a 37 times geometric-mean speedup over the libraries it replaces. The package weighs 26 KB minified with zero dependencies, and existing shadcn projects can switch over with a single CLI migration command.

Aura is a production-tested site-reliability agent platform from Mezmo, built in Rust to investigate and fix incidents in a team's infrastructure. In a demo, three specialist agents correlated traces, latency data and Kubernetes deployment history to trace a payment failure to a regression in productcatalogservice 1.13.2 and recommended a rollback. It runs on OpenAI, Anthropic, Bedrock, Gemini, Ollama or OpenRouter models, keeps the agent system in reviewable TOML, and gates sensitive tool calls behind human approval.

Exo is an AI agent harness, similar to OpenClaw, Pi or Hermes, with full visibility into its own code and logs. That visibility lets it modify its own prompts, memory, tooling and harness policy, clone itself, and manage a lineage of clones, while an unmodifiable event log stops it looping. It runs inside a snapshotable Ubuntu sandbox, needs an OpenAI or OpenRouter API key to start, and cannot yet do generalised computer use of a windowed system.