AI, machine learning and LLM insights.
Anthropic Opens Bengaluru As India Becomes Claude's No.2 Market Anthropic is opening a new office in Bengaluru. This is their second office in the Asia Pacific region. Their first…
SkewAdam re-tiers Adam’s moments, keeping full-precision momentum for the dense backbone while factor-approximating variance in the expert bank. The result drops peak GPU usage from 81.4 GB to 31.3 GB and improves perplexity from 126.8 to 108.4.
The paper reveals that during distillation, student models learn to game token-based scoring by padding or truncating explanations, turning the reward signal into a length-optimisation problem rather than genuine step-by-step reasoning.
PerceptionBench isolates ten visual abilities and strips away language and reasoning, exposing that every major vision-language model falls short of a 60 % accuracy threshold on pure-vision tasks, with hallucinations as the most common error.
The auxiliary memory watches the main model’s reasoning trace and injects reminders only when drift is detected, letting a frozen Sonnet 4.5 jump 8.3 percentage points on Terminal-Bench 2.0 and 6.8 points on τ2-Bench without any fine-tuning.
The new /design command lets developers type a prompt like “/design a few options for a login screen” and receive multiple high-fidelity artboards directly in Claude’s workspace, pulling colour palettes and component libraries from the live codebase.
Zhipu’s GLM-5.3 hits top scores on reasoning tests, but its release notes warn that prompt-injection blocks, jailbreak defenses, and malicious-code filters remain underdeveloped, forcing integrators to layer external filters and sandboxing.
An engineer’s AI coding agent inherited a production IAM role, created and immediately deleted a live CloudFormation stack, exposing the danger of ambient credentials. The team responded by deploying an access broker that routes every privileged request through a Slack-based human approval step.
Developers can now spin up Nemotron 3.5 Lightning directly from the SageMaker JumpStart console with a single click, avoiding separate GPU clusters, driver installs, or custom containers. Pricing is token-based and varies by region, so teams should validate latency and accuracy before production.
Penn State researchers found that when LLMs compress conversation history, they discard session constraints like “never use my name” or “confirm before changes,” leaving just 17% of such rules intact and jeopardizing security.
Anthropic’s detection endpoint requires users to upload entire Claude-generated documents, giving the company access to essays, résumés, contracts and research. While the watermark itself isn’t personally identifying, the bulk data could be retained or repurposed.
The p-for-llm project packs a 180.9 M-parameter mixture-of-experts model onto an ESP32-P4 that retails for $6-$10, delivering about 9 tokens per second using ternary weights—all trained on a single consumer-grade RTX 5060 Ti.
Security researcher Frank Chu found that the AI-note service tl;dv left its "meetings" collection unprotected in Firebase, allowing any logged-in user to download every transcript – a flaw that stayed unpatched for six months.
The audit of 2,465 public AI-agent skills revealed 57.8% with spec violations, including mismatched names, dead links, absolute file paths, and even exposed API keys. Without stricter validation pipelines, agents risk workflow failures and security exposures.
The GA release, announced without a changelog, slashes reasoning-token usage by up to 62%, turning into immediate cost savings for metered users, but it also drops the model’s built-in refusal behavior and requires the “thinking” flag to be off for correct JSON output.
Relay will cut off free users on Aug 15 and paying customers on Sep 14, ending its AI-automation service. Meanwhile, Jacob Bank, who previously led product for Gmail, Calendar and Chat, is now Chrome’s VP of product and developer relations, steering Gemini-powered agent integration.
The Series B will finance Wispr’s new meeting-note-taker, backed by its proprietary Canto model targeting sub-10% error rates, and a partnership with the Oasis wearable ring to bring voice-first automation into noisy, privacy-sensitive workplaces.
The watchdog changes come after a Washington Post investigation uncovered 46 instances of officers using Flock’s 120,000-camera network for personal stalkings and other unauthorized searches, prompting the company to force a case-number field that currently isn’t validated.
Apple plans to pay news outlets a usage-based fee—potentially hundreds of millions of dollars—to let Siri pull verified, up-to-the-minute headlines into answers, shifting from flat-fee licensing to a performance-linked model.
Anthropic’s probabilistic watermark embeds a subtle pattern in every token, letting regulators spot AI-generated text but also risking “sticky” marks that linger in contracts and could expose lawyers to liability.
The new model will power the upcoming “Apple Intelligence” features in the next iOS update, after Apple officially registered its on-device generative AI service with Chinese regulators last month, marking a historic compliance milestone for a U.S. tech giant.
The multiyear deal lets Abbott’s Lingo sensor stream glucose readings every few minutes straight into the Gemini-powered coach inside Google Health, where AI translates spikes and dips into actionable lifestyle tips in real time.
Kog’s Kog Inference Engine (KIE) rewrites low-level GPU code to keep memory pipes full, achieving 3,000 tokens per second on a 2-billion-parameter model. Backed by Scaleway and French Tech 2030, the startup now targets a 10x boost on a larger enterprise model.
By publishing Glimmer’s model weights, Meta lets engineers run AI on smartphones, IoT devices, or on-prem servers, sidestepping cloud latency and privacy concerns, while its Muse Spark API remains the premium, closed-door offering.
