Topic
Self-published evaluation tables and the difficulty of verifying frontier-parity comparisons.
Amazon's new EKS managed addon reads Prometheus metrics from every model pod and sends each request to the one with room in its KV cache. The advertised 82% cut in first-token latency rests on a single 4.4-second baseline.
Reality
TypeSafe's first model, Jev, answers structured questions with typed output and a confidence measure attached. The account of its launch says developers have to check those scores against real outcomes on their own data first.
Reality
The one measured number behind NVIDIA's tokens-per-megawatt pitch comes from Lambda's Blackwell cluster, where power reclaimed from static provisioning ran three extra nodes. Amazon's Annapurna Labs and d-Matrix got a line each.
Reality
Faraj Aalaei's Cognichip argues that semiconductor physics belongs inside the model, with agents left to orchestrate. The one speedup it has documented in detail is a 55-page spec handled in days against a baseline Cognichip supplied itself.
Reality
Astra scored 98.6% on ARC-AGI-3 where its predecessor managed 7.8%. OpenAI is still rationing access while it scales capacity, and that tells a clerical-automation budget more than the benchmark does.
Reality
Ritchie Vink is selling 2.0 as a boring upgrade. Mostly it is, but engine="auto" now resolves to streaming by default, so joins and group_bys stop carrying incidental row order, and pipelines that leaned on it will drift quietly.
Reality
Gemini 3.8 Flash arrived three weeks after Google's last model, its cybersecurity sibling is invite-only, and Google says higher effort levels can spend more tokens. Standardising on a model name now has a shelf life.
Reality
The CameraJet spots a gap between two teeth and fires rinse into it inside 100 milliseconds, which leaves the only proof that the machine learning worked sitting inside a phone app the toothbrush does not need.
Reality
Mercury reports over 1,000 tokens per second per user, and Nemotron Diffusion claims 2 to 8 times autoregressive throughput. Both are vendor figures. The sampler that produces them charges for its speed in arithmetic.
Reality
The zettaFLOPS figure is an estimate. The measured one, sitting in a GB300 example, is a 25 percent density gain, and it is paid for out of the safety margin.
Reality
One SKU, one compute die, 1.2 TB/s of LPDDR5X, and benchmark wins over a 96-core EPYC. The comparison with AMD's Venice is being set on Nvidia's terms rather than on core counts.
Reality
Qwen 3.8 27B ran on a MacBook Pro from a 17GB GGUF and spent 21 minutes on one SVG. Licensing and access stopped being the blocker; latency and KV cache budgeting became the job.
Reality