News
llama. cpp adds Multi-Token Prediction and doubles Qwen3. 6 27 B throughput for local inference
5+ hour, 22+ min ago (478+ words) Ai | llama. cpp's merge of Multi-Token Prediction unlocks large local inference speedups for Qwen3. 6 27 B, with community benchmarks showing roughly 2. 4" on Strix Halo and 2. 17" on other rigs, making on-device open-weight models far more practical for low-latency and privacy-sensitive workloads. llama. cpp…...
Gemini 2. 5 Flash adds a new twist to the AI math race
7+ hour, 41+ min ago (470+ words) Ai | Gemini 2. 5 Flash is being talked about as one of the few frontier models to solve an IMO Problem 6, a milestone that signals how fast AI reasoning is advancing. Gemini 2. 5 Flash has joined the small group of frontier models that…...
Nous Research's new training method could change the economics of LLMs " Startup Fortune
2+ day, 13+ hour ago (553+ words) Nous Research's new training method could change the economics of LLMs Startup Fortune...
Intern LM is making scientific AI smaller with Intern-S2-Preview " Startup Fortune
3+ day, 15+ hour ago (539+ words) Intern LM is making scientific AI smaller with Intern-S2-Preview Startup Fortune Intern LM's new Intern-S2-Preview puts a sharp question in front of scientific AI builders: do useful lab agents need trillion-scale models, or better specialist training?...
AI hacking is turning De Fi security into a balance sheet risk " Startup Fortune
3+ day, 17+ hour ago (216+ words) Ai / Crypto | Google's latest AI-assisted zero-day warning lands at a difficult moment for De Fi after the $292 million Kelp DAO bridge exploit. The issue is no longer only smart-contract code, but whether bridges, verifiers, audits and insurance can keep up…...
Adaption launches Auto Scientist to make model training more adaptive " Startup Fortune
5+ day, 12+ hour ago (607+ words) Ai / Startup | Adaption Labs has launched Auto Scientist, an AI tool designed to automate fine-tuning by improving training data and models together. The product challenges the scale-first economics of frontier AI, but customers will need disciplined evaluations to verify task-specific…...
GPT-5. 5 is turning AI benchmarks into an audit problem " Startup Fortune
6+ day, 16+ hour ago (364+ words) Ai / Business | Epoch AI says an AI-assisted review flagged fatal errors in about one-third of Frontier Math Tiers 1-4. The finding puts fresh pressure on AI benchmarks that investors, founders and labs use to judge model progress. Epoch AI says an…...
Apply Pilot shows how AI job agents are testing hiring systems " Startup Fortune
1+ week, 4+ hour ago (684+ words) Apply Pilot shows how AI job agents are testing hiring systems Startup Fortune...
Meta's email mishap shows why AI agents need real controls " Startup Fortune
1+ week, 1+ day ago (363+ words) Ai / Business | A Meta AI alignment director reportedly lost more than 200 emails after Open Claw acted beyond its instructions. The incident shows why startups adopting autonomous agents need permissions, audit logs, recovery tools and mobile kill switches before handing over…...
Phantom turns five as crypto wallets become crypto's consumer front door " Startup Fortune
1+ week, 2+ day ago (950+ words) Crypto / Startup | Phantom's fifth year shows how crypto wallets are becoming consumer distribution channels for swaps, payments, token launches and app discovery. The bigger question is whether wallet UX and security will decide which chains and apps reach mainstream users....