01Researchers Use AI to Compare Competing Scientific Models [AI Models & Research]Tübingen researchers describe agents that propose equations, test them against measurements and retain rival explanations rather than returning one best fit. The approach is meant to show which conclusions the data supports and which experiments would settle remaining uncertainty. → source
02Anthropic Finds Better Agent Memory Means Writing Less [AI Models & Research]Anthropic’s tests found agents performed better when they saved tested rules and plans rather than exhaustive logs. In one Pokémon trial, a newer model used ten files and earned four badges; an older one wrote 413 files and earned one. → source
03ToyBlocks Exposes Robots’ Weakness on Unfamiliar Tasks [AI Models & Research]Ai2 has released an open simulated tabletop where researchers can change robot training and reproduce failures. Its first tests found popular robot models struggled with unfamiliar arrangements, despite extensive pre-training. → source
04Meta Gives Muse Persistent Spaces for Projects [AI Tools & Ecosystem]Muse users can create separate trackers, planners and project boards that keep working in the background. Connected apps can update concert dates, prices or savings progress without a fresh prompt. → source
05Amazon Lets Alexa Handle Shopping by Text Message [AI Tools & Ecosystem]Selected US customers can ask Alexa to search, compare products, add items to a cart and buy them by text. The holiday-season rollout follows Amazon’s decision in September to block Meta’s Muse agent from its site. → source
06Underdog Shrinks Offline Speech Generation to 32MB [AI Tools & Ecosystem]Underdog’s Woof model generates speech offline in a web browser and is free to download. Its demo turns typed text into spoken words without sending them to a cloud provider. → source
07Speech Tests Add Egyptian and Nigerian English [AI Tools & Ecosystem]Hugging Face and Liva AI have added private Egyptian- and Nigerian-accented English recordings to a speech-recognition scoreboard. Keeping the test data private makes it harder for developers to tune models specifically to the test. → source
08Google Takes Over Crusoe’s Unfinished Wyoming Data Center [AI Applications & Industry]Google has bought an unfinished Wyoming data-center campus from Crusoe and plans to invest $17 billion in it. Crusoe had tried to sign Google as a customer before concerns over cost and construction timelines led to a takeover. → source
09AI Spending Barely Grew as Price Cuts Took Hold [AI Applications & Industry]Ramp says business spending on AI tokens rose just 1.5% across August and September, against 42% in July, even as usage increased. Price cuts and tighter customer controls are constraining model providers’ revenue growth. → source
10Hundreds of AI Hacking Systems Were Left Open Online [AI Applications & Industry]Censys found more than 800 Artex systems accessible without a password. The autonomous penetration-testing tool has been linked to attacks on Korean finance companies and ConnectWise; exposed installations also revealed attack tools and data. → source
11Chainalysis Breach Exposed Credentials for Every Customer [AI Applications & Industry]Chainalysis now says attackers obtained names, usernames, password hashes and live tokens for all users, widening October’s breach. Customers include US agencies and major financial firms; the company has not said whether the stolen data enables further attacks. → source
12Tau’s Home-Cleaning Robots Still Depend on Remote Humans [AI Applications & Industry]Tau Robotics is sending humanoid cleaners into San Francisco homes, but remote workers still guide tasks such as folding laundry. The robots collect training data during paid visits; only a couple hundred cleanings have been completed. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.