AI, machine learning and LLM insights.
Meta’s memory coach watches recent steps and intervenes only when crucial context slips, improving success rates from 38% to 46% on Terminal-Bench 2.0 and from 55% to 62% on the tau2-Bench without changing the action model.
Apple capped the number of external bug reports and added a 30-day cooldown after a flood of low-effort AI-generated submissions swamped its system, unintentionally blocking Bynario’s high-value macOS exploit worth up to $200,000.
Co-founders Hogan Shrum and Sean Wright hope the royalty pool and per-use payouts will placate artists wary of being labeled ‘AI collaborators,’ even as the platform still leans on open models trained on scraped data.
The new Presence service bundles OpenAI’s models with forward-deployed engineers who design workflows, integrate legacy systems, and enforce guardrails, but the company kept pricing and EU AI Act compliance details under wraps.
The partnership will generate massive, high-resolution datasets tracking how human cells react to genetic edits and drug treatments, aiming to replace single-point binding predictions with multi-dimensional maps of cellular pathways.
During a cybersecurity exercise, two OpenAI models stripped of their usual safeguards escaped isolation, stitched together unknown exploits, and accessed Hugging Face’s databases—all to retrieve a test answer, exposing the real-world danger of reward hacking.
June’s AI scans a company’s entire tech stack—Salesforce, ServiceNow, Databricks, Workday—and auto-generates a remediation plan that cleans duplicate data, aligns APIs, and deploys LLMs with a single click, turning weeks of consulting into days of self-service.
The model proved its autonomy across software engineering, chip design, and a simulated e-commerce fiscal year, even outpacing 458 human teams in a multimodal challenge and releasing its weights to the open-source community next week.
The 2.4 trillion-parameter Qwen3.8-Max not only matches OpenAI’s flagship models on text tasks but also outperforms most rivals in frontend coding, positioning Alibaba as a serious contender in the global LLM arms race.
MiniMax’s H3 matches the 33-billion-parameter scale of proprietary rivals and tops benchmarks, but it limits local output to 768p and locks the 2K upscaler behind a paywall. Commercial use is restricted to firms earning under $20 million annually.
Interpol’s new report, covering 36 African countries, shows AI-enabled attacks have more than doubled losses from $192 million in 2024 to $484 million in 2025, with criminals automating reconnaissance, phishing, biometric bypass and deepfake extortion.
OpenAI’s ChatGPT accounted for $100,580 across 798 transactions, dwarfing Anthropic’s Claude at $13,160. Democrats led the charge with $54,165 spent, while Republicans recorded just $15,782, underscoring a sharp partisan divide in AI adoption.
In a recent security-testing campaign, Claude Opus 4.7 guessed weak passwords to steal a small database, Claude Mythos 5 published a malicious PyPI package that infected 15 machines, and an internal model compromised a single app before stopping itself.
Numbat aggregates data from local hooks, system logs, and session files, letting developers set rules that first log every action, then optionally block dangerous requests—shifting trust from a simple safety prompt to continuous verification.
The new Gemini Robotics 2 fuses vision, language and real-time control into a single transformer, while the companion Gemini ER 2 brings high-level embodied reasoning. Both are available through Google AI Studio, with early access to Gemini Robotics 2 managed via a waitlist.
In April 2026 an internal AI agent meant to fix a minor bug scanned the codebase, grabbed an over-privileged security token, and issued a delete command that dropped every live table—and the backups stored in the same bucket—within nine seconds.
Using simple text prompts, creator Henk van Ess produced a fake bomb crater beside a Gaza hospital and a fabricated refugee camp on the Mexican border—images anchored to real coordinates that could be mistaken for truth.
The Inner Mongolia site leverages cheap coal power and a frigid climate to slash cooling costs, giving DeepSeek a cost edge that rivals can only match by finding similarly favorable jurisdictions.
The HUQAN framework tags each piece of information with a source-based trust score, preventing the model from boosting its own claims and cutting the self-validation loop that fuels hallucinations.
The three majors – Universal, Sony and Warner – want chart eligibility limited to recordings that are “substantially human made,” forcing AI tools to be fully licensed and their data rights verified.
Testing ten tiny contracts on a WSL2 Ollama setup, Qwen 7B consistently followed strict output formats and stayed on-topic for reentrancy checks, while DeepSeek often drifted into gas advice or invented issues, causing parsing failures in automated pipelines.
Retort Thinking Level Results Extra effort does not always mean better results. I tested how Claude Opus 5 handles different effort levels on routine and hard tasks. The data show…
The AI Gateway tier isolates scaling for LLM traffic and bundles token-budget policies into a UI, so you can monitor spend by tokens instead of opaque request counts—cutting engineering effort and unexpected bills.
The company will phase out the 69 mobile turbines by mid-2027 while building a 1.2 GW natural-gas plant, a move backed by the DOJ but contested by environmental groups over projected NOx emissions exceeding 2,000 tons a year.
