AI, machine learning and LLM insights.
Nvidia’s new Nemotron 3.5 Lightning runs a 30 B MoE model with only 3 B active parameters, but its NeMo Switchyard routing layer can discard cached prompts when swapping models, effectively doubling the cost of a single inference request.
Tailscale’s engineers traced vanished configuration entries to a 2010 flaw in SQLite’s Write-Ahead Logging. The bug, triggered by high-concurrency writes on certain file systems, silently corrupted databases until patched in SQLite 3.46.1.
Muse Glimmer’s 30-billion-parameter model is quantized to about 17 GB, fitting on GPUs with 24 GB VRAM or Apple Silicon Macs via the llama.cpp runtime. While it shines for on-device agents, its safety tests show higher violation rates than rivals, so developers should add extra safeguards.
FINRA’s Regulatory Notice 24-09 clarifies that every existing supervisory, communication, privacy, record-keeping and best-interest rule now applies to generative-AI use, and examiners will demand inventories, prompt logs and human-review checkpoints.
The revised Directive (EU) 2024/2853 defines any software—downloaded, embedded or cloud-based—as a product, imposing strict liability on manufacturers, importers, component providers and even fine-tuners. Companies must audit code, document training data and update insurance before the Dec 9 2026 de
Hyperdrive sits between Workers and PostgreSQL, reusing persistent connections at the edge so each request avoids the costly authentication round-trips, shaving tens of milliseconds off AI-driven vector queries.
The beta header compresses Claude’s tool-call output, shaving ~14% off those tokens, but because output tokens are a fraction of the total, most users see only a ~3% reduction in their overall API bill, plus a slight latency gain.
Prompt Flow will stop receiving new features on 20 April 2026 and be retired a year later, leaving its runtime images without security patches. Microsoft urges developers to shift to the Agent Framework before compliance issues arise.
Twitch’s chief product officer said an opt-in model would get zero participation, so the platform now defaults to inclusion. Creators can only escape by navigating a hidden toggle in Channel Settings, and Twitch can’t confirm whether past streams are already in Amazon’s models.
Led by Takashi Ishiuchi and Shogo Matoba, the University of Yamanashi team combined somatic-cell nuclear transfer with the Y-CUT CRISPR edit to delete the Y chromosome, producing fertile XO females—even from cryopreserved male cells.
Cognition’s latest financing talks aim to lift its valuation from $26B to $40B, leveraging Devin’s rapid uptake that has driven annualized revenue from $492M to a projected $1B, with enterprise clients like NASA and Goldman Sachs already on board.
The default-on toggle lives in Twitch’s Security and Privacy tab; flipping it stops all live streams, VODs, clips, chat and uploads from feeding Amazon’s generative-AI pipelines while keeping on-platform moderation and recommendation tools active.
Sandbar’s Stream ring uses a deliberate push-to-talk button, sidestepping privacy concerns that have doomed many always-on wearables, while its fully custom hardware—backed by $36 million total funding—targets professional and social settings alike.
A survey of 215 Society of Breast Imaging members shows that while 50% already use FDA-cleared AI, only 35% see a drop in recall rates—far below the 59% they hoped for—highlighting a stark mismatch between expectations and real-world performance.
Anthropic’s Claude Code will stop prompting users to click “approve” for every tool call on August 14, replacing the manual gate with an AI-driven risk classifier that interrupts only on irreversible, destructive, or outward-facing actions.
After weeks of denying AI involvement, D’Addario edited its Instagram post to note Suno Studio had regenerated the demo track, while also admitting it had removed critical comments and blocked users.
The round was led by Menlo Ventures and the Scaleup Europe Fund, with Regent among a dozen other investors. Lovable hit a $500 million annualized run-rate and now runs 60 million projects on a vertically integrated AI stack.
The bug surfaces only when a subclass of Pydantic’s Secret is placed inside a raw dict or list and the endpoint lacks a response model, causing the underlying value to be sent unmasked to clients.
Minute Media has terminated its September 2025 $250 million purchase of Indian AI firm VideoVerse, citing forged merger signatures and missing cash. The move sparked lawsuits from Bluestone Capital, Lingotto and a former COO, all alleging that founder Vinayak Shrivastav siphoned tens of millions thr
In a study with fifteen professional actors portraying patients across five organ systems, Google’s AMIE (Video) earned diagnostic, questioning, and empathy scores on par with those typically given to primary-care physicians, marking the first time a video-based LLM has hit doctor-level parity in a
The Made by Google 2026 event also unveiled Live Transcribe’s camera-based ASL translation, Circle to Search visual queries, and a rugged Pro line with a 40 % thinner camera bar and anti-scratch coating, while the Pixel Tag and Watch 5 expand Google’s hardware portfolio.
Anthropic hired Robert Mahari, a PhD in legal AI and founder of Akiva AI, as its first Head of Claude for Legal. He will steer 12 Claude plugins and more than 20 partnerships—including LexisNexis and Relativity—to embed the model directly into law-firm workflows and curb hallucinations.
Anthropic ran Claude Code through two intensive sessions, deploying 60 sub-agents that issued over 2,400 shell commands and generated 31 million tokens, raising the proven lower bound of zeros on the critical line from 41.6 % to 67.2 %—still a bound, not a proof.
The August 11 rollout adds Copilot Memory for persistent context, native Ollama integration for on-prem inference, and Codex session receipts that log every suggestion—letting teams tune privacy, latency, and auditability per project.
