-
release · Hugging Face
Welcome Beam: a 501B open model you can own, customize, and scale
Hugging Face introduced Beam, an open model with 501B parameters. According to the title, it is a model you can own, customize, and scale.
Because Beam is a 501B open model you can own and customize, teams weighing self-hosting and fine-tuning against API dependence can compare their own serving setup and GPU costs.
-
industry · Simon Willison
Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO
According to the FT, Anthropic told investors that revenue was about $4.8bn in Q1, $11.5bn in Q2 and $16.8bn in Q3. In Q3 it earned $1.70 in revenue for every $1 spent on computing power, with adjusted operating margins of 11% excluding stock-based compensation.
Anthropic's crackdown on distillation, where rivals train models on its outputs, means developers who feed model outputs into other models' training risk terms-of-service violations or restricted API access.
-
research · Hacker News · 2건Korean coverage · GeekNews
Show HN: Gemini 4 Argon plays StarCraft
StarSkirmish Bench v0.2 has LLMs write a Protoss bot for StarCraft: Brood War in C++ within one hour, testing long-horizon reasoning and agentic coding that uses information from practice matches. GPT-6 Astra ranked first with a 57% average win rate, and GPT-6.1 Sol was second at 46%.
It gives developers a benchmark for comparing models on agentic coding that iterates on feedback under a time limit, rather than on single-shot code generation alone.
-
other · Hacker News
Claude reveals ECDSA curve origins
A Hacker News post is titled "Claude reveals ECDSA curve origins." The title indicates that Claude surfaced information about where ECDSA curves come from.
Developers who ask an LLM about the provenance of ECDSA curve choices should check its answer against standards documents before relying on it in security design.
-
research · Hacker News
90% of US working hours are spent on tasks not covered well by AI benchmarks
A Hacker News post reports that 90% of US working hours are spent on tasks not covered well by AI benchmarks. The title suggests existing benchmarks miss most of the work people actually do.
Picking a model by benchmark score alone may not match performance on your real workload, so build your own task-specific eval set before adopting it.
-
industry · Hacker News
Anthropic assigned highest junk credit rating in June
A Hacker News story reports that Anthropic was assigned the highest junk credit rating in June. The headline alone gives no further details.
Developers who build core features on the Anthropic API may want to add vendor financial health to their risk review and check fallback models and multi-vendor designs.