News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedApple’s updated MLX distributes a quantized mixture-of-experts model over Thunderbolt 5 using RDMA. The setup enables private desktop inference, but the largest model still needs several high-memory machines.
Separate complaints accuse OpenAI of misleading customers, failing to protect vulnerable users and collecting minors’ data unlawfully. OpenAI says ChatGPT has age-specific protections and that no system is perfect.
The startup says its bots have handled 2.5 million conversations and collected bank accounts, wallets, phone numbers and malicious links for customers. Its evidence is company-reported rather than independently audited.
Researchers rebuilt two kinds of protein-cutting enzymes around their catalytic cores and tuned them for chosen targets. The designs worked in laboratory tests, though early versions were much weaker than natural enzymes.
Usage data from a gateway that has processed more than 40 trillion tokens suggests developers progress from suggestions to autonomous workflows only when context, routing and feedback remain reliable.
TheoPhysEval covers verifiable, routine research tasks rather than major open questions. Astra scored 98% with web access and 92% without it, but the benchmark does not test whether an agent can choose important problems.
“Trust in Depth” separates job scope, identity, runtime limits, observation and formal proof. AWS says enforceable policy should sit outside the agent, with people handling rules that cannot be stated precisely.
Nemotron 3.5 Super VL and a 24 GB Bolt model are available now, while a 4 GB version is due October 23. Nvidia targets document agents in data centers and local assistants on PCs and edge devices.
An MIT study tested problems from papers published after the models’ training cutoff and blocked web access. Results improved sharply, but the test covered fields that did not require experiments or simulations.
Speakers from Two Sigma, Anthropic and Bloomberg argued that people should review ambiguous or consequential actions, while identity, tool limits, audit logs and kill switches enforce routine boundaries.
The open image model can generate or edit from multiple references and place objects using bounding boxes written into a prompt. Two checkpoints, quantized weights and LoRA training scripts are available.
LangChain recommends starting with 10–20 real traces, testing one behavior and saving a baseline. Production failures should feed the offline suite, while code checks and human-reviewed model judges share grading.
Antigravity filtered millions of uncharacterized proteins and assembled structural, genomic and literature evidence for a 97-amino-acid candidate. Google says wet-lab validation is now underway.
A pilot covering about 1% of Cloudflare’s network gave OpenAI signals for finding useful pages and avoiding broken or unchanged ones. Cloudflare says access permissions and crawler controls still apply.
Timestral comes in four sizes from 5 million to 209 million parameters and handles multivariate, probabilistic forecasts. Mistral says every version runs without a GPU and can be adapted to private data.
Researchers trained a tri-modal masked diffusion model from scratch on text, image-text and audio-text data. The 3-billion-parameter study covers 6.4 trillion tokens and proposes simpler batch-size tuning.
Enhanced protection will use Gemini to assess conversations with unknown senders while local AI checks links before opening. Google says cloud analysis is encrypted and isolated from the company.
The startup’s agents build and refresh company and contact records from more than 100 data sources. An open-source release is targeted for early 2027 after security reviews of the sensitive-data system.
Anthropic distributed model inference across cheap Android boxes and tested agents on varied hardware. No small model completed an end-to-end infection, and the setup relied on exposed debugging access.
A cloud architect can plan tasks while local agents inspect and change source code. Google says one demo kept 97.2% of tokens on-device, though hardware cost, speed and patch quality need separate measurement.
The experimental Earth AI system plans multi-step spatial analyses and can train local prediction models. Partner trials sped up outbreak and disaster work, but benchmark outputs remain drafts for experts.
Artemis Security’s private-preview model links events across corporate systems and can recommend or launch a response. It beat general models on the company’s benchmark, with human approval for high-impact actions.
The open-source method adjusts systematic errors in existing physics and AI forecasts two to six weeks ahead. In one study it improved 98% of precipitation targets and 92% of temperature targets.
An optional filter due before year-end will remove detected AI tracks, albums and related artists from search results. Deezer says synthetic music now makes up more than half of its daily uploads.
White Circle’s RISE method evolved prompts against three image models using ordinary downloadable tools. Human reviewers confirmed bypass rates from about 3% to 13%, and newer systems were reportedly weaker.
The reference system classifies, extracts and validates new document layouts, then saves corrected instructions for later files. AWS estimates about $58 a month for 1,010 sample documents, excluding variation.
Projects across Europe combine video, sensors and digital twins to support city operations while keeping data local. Most remain pilots, though one French deployment reports a 10% operating-cost reduction.
Tech Against Terrorism tested 134 models with more than 200 dangerous prompts and found only two passed completely. Modified open models now appear online in under three days on average.
A lightweight method reads language-selective dimensions inside a pretrained speech encoder to find utterances that switch languages. Whisper-small reached a mean F1 score of 0.798 across 14 language pairs.
Content Understanding adds field-level grounding and a separate confidence model to direct reviewers toward likely errors. One insurer cut the review burden nearly 50%, but still checked every field.
Private Processing would run demanding glasses features inside hardware-isolated cloud machines, separating identity from requests. The architecture is ready for scrutiny, but Meta has not set a rollout date.
Dyna Robotics says its Dyna-2.1 system folded restaurant napkins for 24 hours at a 99.4% success rate. Its world-action model learns from human video, but hardware and training data remain bottlenecks.
HiRO-ACEv2 turns coarse climate simulations into detailed rain, wind and pressure fields at roughly one three-hundredth of X-SHiELD’s cost. It still cannot model historical or warmer climates.
The agentic reviewer checks complex documents section by section and synthesizes inconsistencies across them. Public access includes two free reviews a week, while heavier users can connect a Gemini API key.
The AI infrastructure company says usage tripled in five months and its sandbox product reached a $300 million run rate. Accel led the funding round at a $15.75 billion post-money valuation.
Facebook’s “Your Algorithm” controls let users name interests in natural language or mark posts as wanted or unwanted, with Meta promising feed and Reels recommendations will update immediately.
Anthropic has restricted access to internal reasoning, added detection classifiers and banned tens of thousands of accounts in response to large-scale attempts to copy Claude’s capabilities.
Amex’s new framework asks customers to approve merchant, budget and product constraints, then gives the agent a payment credential tied to that intent instead of the actual card number.
US adults can now launch personal Android apps from Google AI Studio through their regular Google account without publishing them in Play Store search or exposing them to other users.
Epic Games is reportedly shifting an upcoming Marvel and Star Wars shooter to an “AI first” workflow, with people supervising generated code and art after delays and staffing problems.