Vertex Foundry says 2nm yields cleared 70% two quarters ahead of plan
The jump gives Vertex room to undercut its two biggest customers on price — right as both start shopping for a second supplier.
Dele Okafor38 min ago
Breaking
Solace AI ships Orbit-3, closes the benchmark gap with a tenth of the training run
The startup's new reasoning model matches frontier scores on graduate-level math and coding while training on 40,000 GPUs for three weeks — not the usual three months. Rivals are already asking what changed.
Priya Nakamura12 min ago6 min read
Signal — liveon air
Wire activity across desks, updated as dispatches land.
The Desks
Kestrel Autonomy's warehouse arms now re-train themselves overnight
A nightly self-supervised pass lets the picking arms adapt to new SKUs without a human labeling a single frame.
Maya Lindqvist1 hr ago
Argus Institute draft would force labs to disclose training compute over 10^25 FLOPs
The proposal is narrower than last year's bill, but three major labs are already lobbying against the reporting cadence.
Rosa Fontaine1 hr ago
Cobalt Labs raises $85M Series B to put inference caching inside the browser
Ridgeline Ventures leads a round that values the four-person team at $600M, betting local caching can cut API bills in half.
Tomas Reyes2 hr ago
Meridian Cloud's postmortem blames a single mis-scoped rollback script
Tuesday's five-hour outage traces to a config push that skipped canary — the same failure mode the company patched in 2024.
Ada Bergström3 hr ago
Fathom Systems open-sources a 70B model tuned for scientific literature review
It trails the top closed models on general chat but beats them on citation accuracy — the metric its academic users actually care about.
Priya Nakamura4 hr ago
Halcyon Silicon's new edge chip runs a 7B model at 40 tokens/sec on four watts
Aimed squarely at on-device assistants, the part ships in developer kits next month with no cloud fallback required.
Dele Okafor5 hr ago
Loom Systems lets its support agents negotiate refunds within a hard dollar cap
Early customers report a 30% drop in escalations — and a new problem: agents that are slightly too good at haggling.
Tomas Reyes6 hr ago
Continuum's radiology model cuts false negatives 18% in a 40-hospital trial
Regulators will want to see the miss cases before sign-off, but early results are the strongest yet for AI-assisted screening.
Rosa Fontaine8 hr ago
128
Dispatches today
340
Sources monitored
4m
Avg. time to publish
6
Desks covered
The Briefing