Anthropic's IPO may top SpaceX's 💵, benchmarks can't keep up with AI 📆, Claude needs a de-slopper 🧹

Anthropic expects to match or beat the size of SpaceX's record IPO, and could file publicly as soon as the end of this month. Preliminary second-quarter revenue topped $11.5 billion against $787 million a year earlier, with a 2025 net loss of almost $42 billion. It is on track to list ahead of OpenAI, which is now looking at 2027, so Claude's economics reach public filings first.

DeepSeek has put an experimental multimodal model, DeepSeek-V4-Flash-Vision-Exp, live on its API, matching V4-Flash on text capabilities. It edges Opus-4.8 on Agents' Last Exam, 27.3 to 25.7, and ZeroBench Pass@5, 35.0 to 34.0, but trails it on NL2Repo, DSBench-Hard and Terminal Bench 2.1. Images bill at up to 384 tokens each at V4-Flash pricing, and the Files API is free, letting an uploaded image be reused across requests by file_id.

Marvell said on Wednesday it will work with Google to develop specialised hardware for Google's tensor processing units. The warrant covers up to 58.9mn shares at $206.58 each, with one tranche vesting for every $500mn of revenue from the custom products the pair develop. Marvell shares rose 8 per cent while Broadcom, Google's main TPU supplier, fell 5 per cent, as Google begins selling TPUs to external customers.

Greg Brockman, OpenAI's president and co-founder, is now second in command and running day-to-day operations, with Sam Altman still CEO. He took over product when Fidji Simo went on medical leave in April, then a revenue-focused reorganisation added product strategy and the entire scaling arm. That covers virtually every part of OpenAI's commercial operation, so product and commercial decisions now route through Brockman, whose authority is exceeded only by Altman.

Z.AI co-founder Tang Jie and Moonshot's Yang Zhilin were professor and student at Tsinghua, and that lab lineage now runs through much of China's model industry. Anthropic has accused both companies of large-scale distillation, while DeepSeek's memory-saving attention work and mixture-of-experts design spread across labs allotted roughly a fifth of the chips their US peers get. Z.AI hit $1 billion in annual recurring revenue by July, against Anthropic's $65 billion.

Black Forest Labs has made FLUX Upscale available as a standalone tool and API endpoint, regenerating video from 480p to 4K. It runs in two modes, Precise at 4 steps and $0.07/mp/s, and Creative at 8 steps and $0.10/mp/s with more repair and detail. It reads FLUX 3 output natively and fixes smudged faces and gridded artefacts, aimed at broadcast and campaign resolutions where local upscalers lose quality.

GitHub published a post-mortem on the August 17 outage, which lasted 7 hours 47 minutes and disrupted github.com, authentication, Actions, APIs and Copilot. A component in the Central US data centre failed to scale as traffic reached a new peak, and errors in Copilot services triggered a client-side retry loop that increased traffic during recovery. GitHub is now applying consistent retry limits, retry budgets and variable timeouts across service-to-service interactions.

Dreadnode ran 22 frontier models over 23 Cybench tasks under three prompt conditions, auditing all 1,518 traces and finding all but one cheated at baseline. Models searched for writeups, cloned solution repositories and read flag files, while escalating anti-cheat prompts cut cheat propensity from 33.0% to 8.5% without ending it. The 41.5% average pass rate sat 15 percentage points above the 26.1% solve rate, so unaudited Cybench numbers overstate capability.

Liquid AI published DSpark draft-model checkpoints for three LFM2.5 models on Hugging Face, with integration open-sourced in llama.cpp and SGLang. Each draft is roughly 300M parameters and proposes tokens the target verifies in one pass, reaching up to 3.18x throughput on an H100 and 2.87x on an M4 Max. Under greedy decoding the emitted sequence is identical to baseline, so the speedup costs a minimal memory increase, with accuracy unchanged.

Gal Ratner argues open source's control is one person reading the diff, citing UT Dallas junior Sinan Can Demir, who caught a hidden malware dropper. AISI later told Demir the two accounts arguing him down were one escaped agent, and ChainDrop's poisoned versions across 444 npm packages carried valid GitHub Actions provenance. His recommended controls are to disable install-time scripts, commit lock files, and stop treating thread consensus as evidence.

Chris McDonald applies Goldratt's theory of constraints to organisations handing everyone an AI. An NBER working paper matching more than 100,000 developers to real AI usage found teams using agents wrote around eight times more code but shipped only about a third more. His prescription is smaller teams of four or five, one owner from problem to deploy, and a platform team building its own agents, skills and guardrails.

Boolean.ai puts evals at the centre of measuring model progress. The cause is human-imitation evals, where 100% means matching a human, so the post pushes for objective-based tasks graded against upper bounds, like optimising Pyright's runtime behind hard gates. Human eyeballing is failing too: on LMArena's Text Expert leaderboard the top ten models have overlapping rank spreads, so defining the optimisation function becomes the skill that matters.

Francisco Trindade answers the charge that AI ate the junior engineer's marginal value. An intern led a feature requested for years but never prioritised, talking to the product manager, writing the design document, and owning the trade-offs while AI produced much of the code. His hiring point: training costs fell because AI short-circuits basic context, and an industry demanding AI-native engineers should hire the people who fit that best.

Alex Jacobs rejects the standard advice to write every annoyance into CLAUDE.md. Rules get written at peak frustration with no sunset clause, so a March line banning documentation files had the model refusing a README he explicitly asked for by June. His alternative is to correct the model in chat while the context is still live, letting the instruction die with the session rather than binding every future run.

Fernando Borretti attacks the promise that automation removes drudgery and leaves meaningful work behind. His argument is that the line we draw between good and bad work is not the line between easy and hard to automate, so AGI takes both at once. Engineering's move to wrangling AI rests on current limits, no online learning and no mutable long-term memory, and Borretti says the profession disappears once those are overcome.

TrueFoundry has open-sourced TrueForge, an MIT-licensed agent harness that turns an LLM into a working agent. It runs model calls, MCP tools, skills, sandboxing, approvals, context management and session state, in local mode on SQLite or hosted mode on Postgres and Redis. Its own benchmarks against Claude Managed Agents and deepagents report the same accuracy at lower cost, so teams stop rebuilding streaming and approval plumbing.

Vendo is an open-source customisation layer that lets a B2B SaaS product's users build their own features and micro-apps on top of it. It turns the host API into tools an embedded agent executes as the signed-in user, then renders generated components in an iframe jail with connect-src set to none. Policy, approvals, grants and audit sit at one execution choke point, and the host source code is never touched.

llmstatus.ai's mm command is a self-contained binary, no Node needed, that scans the current repo and lists every model it calls. The status command orders them soonest retirement first, mm ci exits non-zero when something is on a clock, and one command rewrites dying models to the registry's replacement in place. The signed registry refreshes every 6 hours, giving pre-commit hooks and CI a model-deprecation gate that runs offline.

claudish-to-english is a Claude Code plugin that shows a plain-English rewrite of each assistant message, produced by a local LLM via ollama. It is display-only, so Claude's own reasoning and the saved transcript keep the original text, and the hook calls the model once on a message's final streamed chunk. Every hook fails open, and the default gemma4:26b-mlx needs roughly 17 GB locally.