GPT-Live goes full-duplex 🔊, Bun rewritten in Rust by Claude 🧑‍💻, OpenAI: SWE-Bench is 30% broken 🧑‍💻

OpenAI released GPT-Live-1 and its smaller mini variant, full-duplex voice models that speak and listen at once, so you can interrupt naturally and get live translation. They replace the old transcribe-then-LLM-then-speak chain, and hand queries to GPT-5.5 for search, reasoning or agentic work while the conversation keeps flowing. GPT-Live-1 mini becomes the default voice in ChatGPT, with the full model reserved for paid tiers.

Grok 4.5 is SpaceXAI's strongest model yet, trained alongside Cursor weeks after SpaceX agreed to buy the coding startup. It handles coding, agentic and office work, tops Harvey's legal-agent benchmark, and says it has roughly twice the token efficiency of rivals. Musk called it Opus-class, though on DeepSWE and SWE-Bench Pro it sits a notch below Fable, GPT-5.5 and Opus 4.8, competitive but not best-in-class.

Cognition, the maker of Devin, trained SWE-1.7, a coding-agent model on a Kimi K2.7 base that reaches near-frontier quality at a fraction of the cost, close to GPT-5.5 and Opus 4.8 on agentic-coding benchmarks. The gains atop an already RL-trained base suggest RL can push past a supposed post-training ceiling. Read the numbers with care, since these benchmarks can overfit to a product's own workflow.

Entire, from GitHub's former CEO Thomas Dohmke, is a distributed Git-compatible network built for the moment coding agents fire thousands of concurrent clones a second and blow through a repo's rate limits. You mirror your GitHub repositories onto regional Entire cells that absorb the heavy read traffic, keeping the origin on GitHub while agents fetch fast nearby. It is live now across the US, EU and Australia, new users waitlisted.

Researchers built HalluSquatting, the first prompt-injection attack that scales to mass botnets, DDoS and device infection. It exploits a pull pattern: LLMs confidently invent package or resource names, so attackers pre-register those hallucinated names and wait for the tools to send victims their way. It works against nine assistants and agents including Cursor, Gemini CLI, GitHub Copilot, Cline and OpenClaw, all susceptible straight out of the box.

Even before OpenAI and Anthropic go public, their paper wealth is warping the San Francisco housing market, with some sellers asking for equity rather than cash. Bay Area home sales rose 12.2 percent year over year in May and the median single-family price hit $2.14 million, as buyers flush with pre-IPO stock race to close. The clearest sign yet of lab valuations spilling into the physical economy.

Meta is prototyping super-sensing smart glasses that continuously record audio and snap photos every few seconds, the Financial Times reports, letting the wearer later ask Meta AI about what they saw. One proposed design keeps the raw footage with Meta rather than the user and leaves the recording light off. It lands against Meta's record of funnelling user media to contractors to train its AI, private moments included.

AI Now Institute researchers published a proof-of-concept that pulls remote code execution out of Claude Code and Codex CLIs when they are pointed at a library to assess its security. The exploit needs only Claude Code's auto-mode or Codex's auto-review, with prompt injections scattered through the library's own source, no hooks or MCP servers required. They warn it undercuts the White House push to accelerate defensive AI.

Jarred Sumner walked through why Bun, the JavaScript runtime he first wrote in Zig, is being rewritten in Rust, much of it with a pre-release Claude Fable 5. Bun now clears 22 million monthly downloads and underpins Claude Code and OpenCode, but its huge scope made memory-safety bugs a constant tax. Rust plus an AI pair is his bet on stability, and Bun is now Anthropic-owned.

Databricks built an internal benchmark that scores coding agents on real edits its engineers made against a multi-million-line production codebase, spanning Python, Go, TypeScript and Scala, rather than synthetic tasks. It plots cost against quality and finds the Pareto frontier, best quality for a given price, runs through both OpenAI and Anthropic models. The team says the exercise has already made its engineers meaningfully faster with agents.

In a GPU Mode auto-research contest, one entrant looped Codex on a batched Householder QR-factorisation kernel and finished 12th of 183 with a 212x speedup over the baseline. His writeup is the method, not the maths: learn the problem well enough to ask better questions, keep the serial work small, and inject idea diversity so the loop escapes local maxima. A concrete playbook for loop-engineering an optimisation problem.

OpenAI ran a detailed audit of SWE-Bench Pro, one of the coding benchmarks labs cite most, and estimates that roughly 30 percent of its tasks are broken by design flaws or contamination. Using human-supervised agent review plus a human annotation campaign, it warns that faulty evals give a false read on capability, distorting safety cases and research priorities. It follows OpenAI's earlier finding that SWE-bench Verified had lost meaningful signal.

Addy Osmani makes the case that as agents take on more of the work, engineers have to own the outer loop, the accountability for what ships. He breaks it into three moves: Quality installs the checks that produce evidence, Verdict is the human decision to ship, block or reject, and Answerability is being able to explain why. An agent writes the code, but someone still has to stand behind it.

Adapted from a talk on vibe coding, these notes lay out the software factory: give the model a developer's toolkit so it checks its own work. That means CLI access to deployments to catch its own bugs, a browser to see the frontend, big compute to run tests, even a funded wallet. Then find the workflow's slowest step and speed it up, often with a monitoring loop.

Every deck says AI startups will gut incumbents and the system of record is dead. Adam Harris, who built Cloudbeds into a hospitality platform doing $20 billion a year across 157 markets, pushes back: last month it displaced 117 rival systems he had never heard of. His moat is 14 years of proprietary pricing data wired into an AI layer, even as 80 percent of incumbents stay too slow to survive.

TypeScript 7 is a native port of the compiler rewritten in Go that runs 8 to 12 times faster on full builds. The team rebuilt the toolset faithfully, keeping the original structure and logic so results stay compatible, while gaining native speed, shared-memory multithreading and new optimisations. It installs the usual way over npm and puts a much faster tsc in your workspace, with editor tooling following.

Microsoft's SkillOpt treats an agent's natural-language skills like something you can train, with epochs, batch sizes and validation gates, but without touching model weights. The new v0.2.0 adds SkillOpt-Sleep, a nightly offline engine that harvests past sessions, mines and replays them, then consolidates the survivors behind a held-out validation gate. It brings cross-tool backends for Claude, Codex, Copilot, Devin and OpenClaw, so gains carry across whichever agent you run.

Morph is an embeddable AI prompt bar you add to your product with one script tag and a key. You register the data you already have, and your users describe what they are missing to get a real working widget computed from that data, with no wait on your roadmap. It also reshapes and restyles your existing UI by prompt, is MIT-licensed, and keeps the data in the page.

Microsoft's Flint, now open source, is a visualisation language that lets an agent describe a chart by field meaning, not pixel-level config. You tag columns with semantic types like Rank, YearMonth or Delta, and Flint infers the parsing, scale, axes, formatting and colour scheme, then compiles a chart spec. The pitch is charts an LLM can reliably generate, since intent maps to output without hand-wiring each encoding.