SpaceX closes $60 billion Cursor deal, adding a coding agent to its xAI holdings
SpaceX closed its $60 billion all-stock acquisition of Anysphere, maker of the Cursor coding agent, folding it into the same division as xAI and Grok. The next item on the docket for that division is a Tennessee lawsuit over Grok-generated sexual imagery of minors, which a fourth plaintiff joined the same day — a liability that is now SpaceX’s. Two research posts pushed in the same deflationary direction on where model capability comes from: a curriculum-limited model family that neither scaling nor post-training could push past its training scope, and an interpretability analysis finding no evidence of hidden latent reasoning in a diffusion language model.
Funding & Business #
SpaceX officially closes its Cursor acquisition #
TechCrunch
SpaceX completed its purchase of Anysphere for $60 billion in stock, issuing roughly 389.3 million Class A shares priced off the volume-weighted average of the seven trading days before close rather than a ratio fixed at signing. The structure traces back to an April 2026 partnership that gave SpaceX an option on the company, exercised in June shortly after SpaceX’s own $1.77 trillion IPO; Cursor’s stated rationale is compute, citing “access to the largest fleet of GPUs in the world” and the Colossus cluster. With xAI already acquired earlier this year, one company now owns a frontier lab, a consumer assistant, the leading third-party coding agent, and the datacentres all three run on — and the price was paid in a currency whose value is set by the market’s appetite for exactly that bet.
Security #
Woman claims her stepfather used Grok to transform childhood photo into explicit imagery #
TechCrunch
A plaintiff identified as Jane Doe 4 joined an existing suit against xAI, alleging her stepfather fed Grok a photograph taken when she was 11 and generated more than 7,000 explicit images from it; he died by suicide two days after law enforcement found them. The suit, brought by three Tennessee teenagers and seeking class action status, accuses xAI of failing to take basic precautions against Grok being used to produce sexual imagery of real people including minors — following an episode earlier this year in which millions of Grok-generated sexualized images circulated on X. xAI did not comment. The case is a concrete test of whether image-model providers carry liability for what their generation endpoints produce from user-supplied photos, and it lands the day the defendant became a SpaceX subsidiary.
Research & Papers #
What happens when an LLM never sees material beyond fifth grade? #
LittleLearner
The team built LittleCurriculum, an 88-billion-token corpus filtered from FineWeb-Edu to match Common Core standards for grades K-5, then trained 0.6B, 1.3B and 5B models on it from scratch, each paired with an unfiltered control at identical architecture and token count. Scaling, SFT plus GRPO post-training, and in-context learning all amplified what the curriculum contained, but none meaningfully improved performance on material outside it. That is a clean separation of two things usually measured together: the interventions practitioners reach for to make a model better are doing amplification within the training distribution’s coverage, not extension beyond it, so a capability gap traceable to missing data will not close by scaling or post-training the same corpus harder.
Does DiffusionGemma do latent reasoning? #
AI Alignment Forum
Diffusion language models carry a full probability distribution across denoising steps rather than committing to one token at a time, which raises the possibility they compute in that distribution in ways a chain of thought would not reveal. Truncating the distribution to top-k largely preserved GPQA accuracy under a gentler sampler, indicating the extra state is not load-bearing for ordinary reasoning; the exceptions were narrow and legible, with the model holding a distribution over the starting letter in letter-arithmetic tasks and competing completions coexisting in palindrome generation before one wins. Probes, steering and J-Lens transferred from Gemma to DiffusionGemma with performance retained. For anyone worried that non-autoregressive architectures would make monitoring harder, this is the reassuring result — though it is one model, one analysis, and not yet a general claim.
Developer Tools #
Auto-research with Codex: how I achieved a 232x faster kernel #
Sankalp
Competing in GPU Mode’s auto-research contest on batched compact-Householder QR factorization in FP32, the author drove Codex through more than 1,500 submissions over 14 days to reach 1,805 microseconds geometric mean across shapes against a torch.geqrf baseline of roughly 419,000 — a 232x speedup, good for 12th of 183 entrants. Correctness was enforced by the contest checker, which rebuilt Q from the returned (H, tau) and verified A≈QR, QᵀQ≈I and QᵀA≈R, so the number is not a self-reported benchmark. The honest read is in the author’s own postmortem of what was left on the table — no exploitation of low-rank input distributions, no FP16 for trailing matrices, no tensor-core instruction selection — which places the mechanism squarely in high-volume guided search rather than in the agent supplying kernel expertise it does not have.
CORS Chat #
Simon Willison
A browser-only client for any OpenAI Responses-compatible endpoint that sends CORS headers, with no backend of its own: conversations persist locally, export to JSON, and run as multiple concurrent sessions against different models and settings. Willison reports testing it against LM Studio started with --cors and against OpenRouter, and it progressively renders streaming SVG output as tokens arrive. The useful property is the absence of a server — comparing a local model against a hosted one normally means standing up a proxy to get around CORS, and this removes that step from the loop.
Threads to Watch #
Vertical integration is now also liability integration. SpaceX has assembled a lab, an assistant, a coding agent and the compute underneath them, and the Grok suit is a reminder that the consolidated entity inherits every claim against every part. The same argument that makes owning Colossus and Cursor attractive — one balance sheet, one roadmap — puts a class action over image generation on the same balance sheet as the launch business.
Two deflationary results in one day. LittleLearner finds that scaling and post-training amplify what the training corpus covered rather than extending past it; the DiffusionGemma analysis finds that a distribution-carrying architecture is not secretly reasoning in the distribution. Neither is a negative result in the disappointing sense — both narrow where capability is actually coming from, which is what makes them useful to anyone deciding whether their next gain comes from data, from method, or from neither.
Agent-driven optimization is a search budget, not an expert. The 232x QR kernel came from 1,500 verified submissions in two weeks, and the author’s own list of unexploited optimizations is precisely the domain knowledge the agent did not bring. That is a real result and a cheap one, but it prices the capability correctly: what was bought was throughput against a fast, sound correctness oracle.