A federal judge just drew a line around who gets to dictate an AI model's behavior—and Silicon Valley's building hardware for the next fight. This is AI Daily Briefing. A court ruling, a chip package, and two research papers that should make a few product roadmaps sweat. We'll start with Anthropic, because the government just found out procurement power has limits. Here's Ars Technica:
The Trump administration's blacklisting of Anthropic was illegal, a federal judge ruled in an order vacating government directives against the use of the firm's AI technology. The government illegally retaliated against Anthropic by designating it a supply-chain risk to national security, said yesterday's ruling by Judge Rita Lin in the US District Court for the Northern District of California.
Judge Rita Lin vacated the directives and called it First Amendment retaliation. To be precise, the administration used a national-security label after Anthropic held the line on lethal autonomous warfare and mass surveillance. For federal procurement teams, that sets a serious new floor. You can't blacklist a vendor because its model-use policy annoys you, then dress it up as supply-chain risk. So that supposedly abstract open-versus-closed fight is now in court: who controls the restrictions and fine-tuning around a model? The government tried to take that control here. It lost. The military-use restrictions are where contract lawyers will stay busy. But Judge Lin’s finding means agencies now have to argue an actual procurement case—not use a content-policy dispute as a kill switch. Ben Wodecki, writing in SDxCentral:
Meta has put networking front and center in its next-generation custom silicon, placing network interface cards (NICs) inside the chip package itself to reduce interconnectivity bottlenecks. Its Meta Training and Inference Accelerator (MTIA) 300 packs two network chiplets – each packing six custom 800 Gb/s remote direct memory access (RDMA) NICs – to provide 1.2 TB/s of total I/O bandwidth without the need to touch a peripheral component interconnect express (PCIe) bus.
Meta put twelve custom 800-gig RDMA NICs directly into MTIA 300—two chiplets, six each—and skips the PCIe hop entirely. That’s a very expensive way to admit network latency was eating the workload alive. And it’s tied to the part of Meta that prints money: ranking and recommendation. Those embedding tables hold more than 99% of the parameters, so moving data fast matters at least as much as headline compute. Meta reports a 3.9x communication speedup on a 150-billion-parameter recommendation workload. It’s a real workload, with a real number tied to a bottleneck you can point to—already more useful than most chip-launch decks. This is the infrastructure story people will skip past for the legal drama. Meta’s $59.4 billion in ad revenue gives it every reason to redesign the whole inference stack around its feeds, and on-package networking is where that advantage starts to compound. This one's from TechNode:
PsiBot, a Chinese embodied-AI startup focused on dexterous manipulation, has completed a new funding round of more than US$100 million. The round includes Ningbo Tuopu Group, a Chery Holdings-backed fund, Lens Technology, 37 Interactive Entertainment, Wuhu Investment Holding Group and Fosun Fortune Capital. Existing shareholder Zhuhai Technology Industry Group also continued to invest.
PsiBot pulled more than $100 million, and the investor list is the signal. Ningbo Tuopu, a Chery-backed fund, and Lens Technology can actually put dexterous manipulation on a factory floor. Plus 37 Interactive Entertainment, because apparently every serious AI cap table needs one name that makes you blink. But logistics and advanced-manufacturing deployments are a much better use of this money than another humanoid demo doing a slow-motion coffee pour. Their Psi-R2 policy model and Psi-W0 world model look tidy on a slide. The expensive part is collecting human-operation data—enough examples of hands, objects, bad lighting, weird packaging, all the stuff that ruins a lab run. And industrial backers can supply those ugly real-world edge cases. That gives this round more substance than the usual embodied-AI money: the capital is tied to places where the robot has to earn its keep. IBM Research writes:
Document processing pipelines traditionally cascade optical character recognition (OCR) engines with downstream models for structured information extraction, leading to multi-stage error propagation. We fine-tune SmolDocling, a compact 256M-parameter vision-language model (VLM), to perform end-to-end key-value extraction directly from document images, jointly solving identification, localization, and association in a single pass without OCR preprocessing. We extend DocTags with specialized key, value, region, and link tags, enabling many-to-many relationships in a unified output sequence.
IBM fine-tuned a 256-million-parameter SmolDocling to pull key-value pairs straight from document images—no OCR handoff, no three-stage pipeline to poison the next stage's output. For invoices, forms, claims—all that grim enterprise paperwork—that's a very practical target. And it beat larger zero-shot VLMs on IBM's layout-aware evaluation while coming in 27 times smaller than Qwen2.5-VL 7B. The evaluation checks bounding boxes as well as text, which matters when the answer is right but attached to the wrong field. Over five times faster at inference is the line procurement teams should circle. A compact model that reads a form in one pass can beat an OCR-plus-general-model stack on cost and latency, with less error cleanup—assuming that private-dataset result survives contact with your ugly scans. Here's Apple:
Our extensive experiments evaluate multiple approaches in which LLMs can incorporate evidence into their beliefs. Some of these approaches produce (nearly) Bayesian updates, thus optimally processing evidence; others use a learned heuristic. Surprisingly, the non-Bayesian heuristic updates often outperform exact Bayesian updates (optimal information processing) in terms of downstream task performance—indicating the LLMs’ probabilistic models of the world are misspecified.
Apple and Stanford tested how models revise probabilities as new evidence arrives, and found the learned, non-Bayesian shortcuts often beat the mathematically correct update. That's a pretty clean warning label for anyone selling an LLM as a reliable inferential reasoner. Right—if the heuristic wins downstream, the model's internal picture of the world is miscalibrated. Don't wire its confidence score into a medical, legal, or fraud workflow and call it uncertainty quantification. And notice the contrast with the IBM document model we just covered: a compact system can excel at a bounded extraction task without pretending it has coherent beliefs about evidence. Apple’s paper gives builders a way to spot where that pretense breaks. If you’re finding AI Daily Briefing useful, please subscribe or leave a review wherever you’re listening. Reviews help other people find the show, and we’re grateful you’re here.
Links to every story we covered are in the show notes if you’d like to dig deeper into anything that caught your attention. Thanks for listening. We’ll be back tomorrow. That’s AI Daily Briefing for today. This is a Lantern Podcast.