Tech News
← Home  ·  All topics

Model

339 GoKawiil briefs on this topic · RSS for model

Researchers seize Hugging Face account for Kokoro text-to-speech model

Security researchers led by Alon Hertz took control of the Hugging Face account for Kokoro, a popular open-weight text-to-speech model, by registering the expired domain listed as the developer's contact and using it to reset the account's password. They then demonstrated that an attacker could quietly alter the model's neural weights so it generates audio diverging from input text, without modifying any application code. Kokoro, abandoned by developer Hexgrad since April 2025, still receives nearly 13 million monthly downloads and is used in tools from Google, NVIDIA, AMD, Intel and AWS.

California Reportedly Sues OpenAI Over AI Model Oversight Issues

California has filed a lawsuit against OpenAI concerning the loss of control over AI models distributed through platforms like Hugging Face, according to reports. Specific details of the legal claims and requested remedies were not included in the available material.

Entrepreneur op-ed: agencies survive AI disruption via flexible contracts, client trust

An agency owner writing for Entrepreneur describes 2025 as the worst year for new client acquisition in their firm's seven-year history, attributing the decline to AI tools letting businesses solve problems in-house. The owner says their agency has avoided layoffs or revenue drops by offering flexible, non-restrictive contracts and actively working to prevent client resentment, rather than locking clients into rigid agreements.

OpenAI signs deals with Apollo, METR, Transluce and Goodfire for safety reviews

OpenAI announced agreements with four independent evaluators—Apollo Research, METR, Transluce and Goodfire—to conduct pilot assessments of its AI models' behavior and safeguards. The company said it will grant these groups deep access to its internal systems and use the results to shape future collaborations with outside safety experts.

Garlic LLM runs Qwen3.5-9B AI model on jailbroken PS4 hardware

Developer Leonardo Alegría released Garlic LLM, an open-source app that lets a jailbroken PS4 run Alibaba's 9-billion-parameter Qwen3.5-9B model entirely offline, using the console's 2013-era GPU. The model is compressed to roughly 4.5GiB and runs at 31-39 tokens per second, with voice chat and dictation features powered by additional speech models. It requires a PS4 running the GoldHEN homebrew enabler and is distributed as a free 5.66GB download via Hugging Face.

New Open-Source Plugin 'Elephant-Goldfish Model' Aims to Improve AI Planning Review

A new open-source plugin called the Elephant-Goldfish Model (EGM) has been released for AI assistants. It lets a user develop a plan with an assistant that has full context, then has separate reviewers—without access to that conversation—read only the resulting plan and flag what is missing or unclear.

New open-source tool Turbo-Browser-MCP cuts browser automation tokens for LLMs

A developer released turbo-browser-mcp, an open-source Playwright-based server implementing Anthropic's Model Context Protocol, aimed at reducing token usage and latency when LLMs automate browser tasks. It merges dozens of granular browser actions into 10 batched tools, adding features like delta DOM snapshots, single-step form filling across React/Shadow DOM/iframes, direct file uploads without Base64 encoding, parallel tabs, network blocking, and a default human-confirmation step before submitting forms. The project's own benchmarks claim token savings up to 99.8% and speed gains up to 7x on common automation tasks.

Tesla opens Cybercab sales to select buyers with mandatory $299 monthly fee

Tesla is now inviting select customers to buy a Cybercab, its steering-wheel-free two-seater, for personal use, but only in cities where its Robotaxi service already operates. The car is priced at $29,990 plus a required $299-a-month 'Autonomy Subscription,' according to screenshots shared by Tesla's Robotaxi account on X. Tesla currently runs 278 Cybercabs, about 40 percent of its Robotaxi fleet, per the Robotaxi Tracker.

Researchers demonstrate model-to-model weight-editing attack that corrupts factual answers

A research paper describes a controlled experiment in which one AI model acted as an 'editor' and directly altered another model's weights, causing it to give a false answer (naming Buenos Aires instead of Sydney as host of the 2000 Olympics). The authors call this 'weight-space misfolding'—a persistent change to a model's parameters that exceeds the authority granted for legitimate editing, distinct from a temporary prompt-based manipulation.The test was conducted on a small model and is not evidence of this occurring in deployed systems or spreading between models in the wild.

uOttawa physicist's model pegs universe's age at 26.7 billion years

Rajendra Gupta and Nikolaos Samaras at the University of Ottawa say their CCC+TL model, which alters assumptions about gravity and light rather than invoking dark matter or dark energy, estimates the universe's age at 26.7 billion years, roughly double the standard 13.8 billion-year figure. The model, published in Monthly Notices of the Royal Astronomical Society, reportedly matches observed hydrogen and helium abundances, historical expansion rates, and the orbital behavior of stars and gas in 175 galaxies without requiring dark matter.

Developers release open-source Edge0-145B on-device AI model

A development team has open-sourced Edge0-145B and Edge0-120B, two on-device AI models, alongside an inference framework and deployment tools. Edge0-145B is described as the first open-source on-device model with 145 billion parameters, running locally on a compatible laptop with reported peak memory of 29GB and decoding speeds above 80 tokens per second.

Google introduces Orchestrator to mix frontier and lightweight models in Antigravity

Google has launched Orchestrator within Google Antigravity, a system that automatically switches between models — codenamed Carbon, Argon and Hydrogen — during a single coding session rather than relying on one model throughout. Hydrogen is positioned as a fast execution model for routine coding tasks, working alongside Carbon and Argon, which handle deeper architectural reasoning. On Terminal-Bench 4.0, Google reports Orchestrator achieved task success rates close to Carbon alone while using 24% less quota and finishing 19% faster.