Commentators Warn It Will Be 'Unsafe' for Humans to Write Code as OpenAI Prepares Its Own Cyber Model Rollout

Following Anthropic's Mythos announcement, commentators warn that "it will be unsafe" for humans to write code at all, given Mythos's superhuman vulnerability discovery.

Mythos appears to be the first model class trained at scale on Blackwells, with Vera Rubins waiting next. Blackwell and Vera Rubin are NVIDIA's latest and upcoming GPU architectures.

OpenAI is reportedly finalizing its own Mythos-style staggered rollout of a cyber model to a small set of partners.

Elon says SpaceXAI's Colossus 2 now has 7 models in training, from Imagine V2 through twin 1T and 1.5T variants up to a 10T model, with each pretraining run lasting roughly two months.

A leaked memo from the post-merger xAI's new president admits the lab is "clearly behind" the other frontier shops and is reorganizing engineering ahead of the SpaceX IPO.

Meta's Muse Spark, the first model under Alexandr Wang, is being called "a data labeling CEO's model" for strong data-quality benchmarks while underperforming on reasoning ones.

Alibaba anonymously dropped HappyHorse-1.0, which took the #1 slot on Artificial Analysis's text-to-video and image-to-video boards, knocking ByteDance's Seedance 2.0 to second.

ByteDance is fighting back with In-Place Test-Time Training, repurposing MLP projection matrices as fast weights so a 4B model can perform well at 128k context. Fast weights are parameters that update during inference rather than just during training.

OpenAI's researchers solved 5 more Erdős problems across combinatorics, probability, and number theory.

Cognition's Scott Wu notes global FLOPs are growing roughly 3x annually while inference demand is growing roughly 10x, forecasting price hikes and a flight to smaller models.

Perplexity's ARR doubled to $500M since New Year's. ARR stands for annual recurring revenue.

Tubi became the first major streamer to launch a native app inside ChatGPT.

Google launched Notebooks in the Gemini app, folding NotebookLM directly into the assistant so chats, sources, and files share one workspace.

Syncere unveiled Lume, a lamp-shaped robot pitched as something that "does your chores."

TSMC's CoWoS packaging is compounding at 80% annually, with the majority of capacity earmarked for Nvidia. CoWoS is an advanced chip packaging technology that connects multiple chips together.

Meta committed an additional $21B to CoreWeave running through 2032, atop a prior $14.2B deal.

OpenAI paused its UK Stargate buildout, citing energy costs and regulation.

Epoch AI calculated Chinese and open labs are running on roughly 10x less compute than the frontier.

Germany is building the world's tallest wind turbine, 364 meters, inside a coal mine.

Life Biosciences raised $80M to begin clinical testing of its anti-aging gene therapy.

GLP-1 drugs are projected to add $13B in apparel sales as Americans shrink out of their wardrobes.

The iPhone Fold is reportedly on track for a September launch.

Rep. Ogles says the White House registered "Aliens.gov" because the President wants to be "the guy that revealed the truth" and lay a "historic" baseline on UAPs.

Rep. Burchett's HR 8197 would dissolve AARO entirely.

UBS's HOLT model, a cash-flow valuation tool, now pegs Nvidia's fair value at $22 trillion.

OpenAI CFO Sarah Friar says retail investors will "for sure" get IPO shares after strong demand from individuals in the latest round.