The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
A new Hugging Face Space shows whether popular Hub models run unmodified on AWS Trainium via torch-neuronx, with cores, wall time, memory, scripts and failures in an open dataset. It is honest, checkable compatibility data; a green check means 32 matching tokens, not production performance.
Anima reports clinic phone waits falling from 40–50 minutes to about 10, document handling doubling or tripling, and backlogs of 100,000+ documents cleared in days across 1,000+ UK clinics. Aiming AI at medicine’s admin queues is sound; the figures are self-reported and coding accuracy needs independent checks.
CNBC reports OpenAI is close to raising up to $30 billion at a $1.25 trillion pre-money valuation, led by Coatue and UAE investors, possibly paired with compute. A month earlier Bloomberg reported $1.4 trillion. Terms aren’t final; the bigger question is why keep raising privately after a confidential IPO filing.
Meta removed training language from its business terms after publisher pushback, but legal experts told Nieman Lab that undefined phrases like ‘develop’ a ‘business product’ and content ‘associated with’ a business could still cover AI use of scraped sites. Meta says robots.txt still governs; it hasn’t written that in.
SpaceX and Tesla have formed special committees to discuss a merger that Bloomberg reports could be announced by mid-2027. The AI logic is pitched as superintelligence, but SpaceX’s AI revenue looks like compute rental, and the clearest gains are Musk’s voting control and a possible pay-package trigger.
1Password’s new service gives each agent a key, asks the user to approve each login, and fills credentials itself so the agent never sees them. It protects the secret, not what the agent does after login, and it lands as banks say they can’t tell customer-requested agents from fraud.
OpenAI’s answer to the fast, cheap decision-model category is a reconfigured ChatGPT Luna. It cuts response times to 100–180 milliseconds but trails rivals on leaderboards, and in this category calibration matters more than speed.
Canva’s first in-house image model, Craft, learned from human-made templates and plans a layout before drawing, aiming at legible text and editable designs. It also cuts reliance on rivals’ APIs. How template creators are credited or paid is unanswered.
In a peer-reviewed Science simulation, Microsoft’s MAI-DxO diagnosed hard NEJM cases correctly 83% of the time versus 36% for generalist doctors. The comparison wasn’t with specialists, most of the gain came from stronger models, and a parallel trial at Beth Israel Deaconess is the real test.
On Meta’s Q3 call, Zuckerberg pitched a network where people’s Muse agents talk and transact with other Muses, predicting Muse will out-earn Facebook and Instagram per user. Early adoption is real; a free agent paid by transaction fees, a shaky trust record and record capex make the guardrails the story.
Liquid AI says its new RIPL engine puts small generative models inside millisecond request paths, and the claim is promising, but every number so far is Liquid's own.
Modder Sammy Hajhamid (pepsipu) says OpenAI’s Codex helped build ‘ns2jb,’ a tethered, software-only Switch 2 jailbreak launched through SmileBASIC 4. The immediate result is Peter Griffin in Mario Kart; the signal is a coding agent speeding exploit work on a locked current-gen device.