Encrypted Reasoning Traces Are Not A Security Boundary
Encrypted Reasoning Traces Are Not A Security Boundary A new research paper shows a major flaw in how AI providers handle reasoning. Anthropic, OpenAI, and Google send encrypted r…
GyaanSetu AI
AI, machine learning and LLM insights.
Anthropic Hides Billions of Debt in SPV, Potentially Raising API Prices
The SPV, called Theseus Infrastructure, will raise equity, borrow private-credit funds, build the data centres and lease them back to Anthropic. Lease payments must cover that borrowing, meaning developers could see higher Claude API rates if the hidden debt costs rise.
Switching to Structure-Aware Chunking Boosts RAG Hit Rate from 61% to 83%
After four months of moving a RAG pipeline out of a notebook, the author shows that changing chunking strategy alone lifted the retrieval hit rate to 83%, while a 200-query evaluation set now catches most regressions before they reach customers.
OpenAI’s Ethics Lead Walks Out, Adding to Wave of Safety Staff Exodus
The resignation follows a string of departures from OpenAI’s safety teams, including the Superalignment group in 2024 and board shake-ups in 2023, underscoring growing friction between rapid product releases and internal risk checks.
Even Top Tattoo AI Misidentifies Cases, Raising False-Positive Fears in Courts
The NIST study forced a rethink of how tattoo-search tools work, urging developers to drop binary matches, show similarity scores, and let trained analysts make final calls before any legal use.
Accel Raises $550 Million India Fund, Aiming to Fuel AI-Built Applications
Accel’s latest $550M vehicle is part of a $3.5B global push into India, targeting AI-enabled fintech, advanced manufacturing and deep-tech startups, with a playbook that favors building applications on existing large-language models.
Subquadratic’s SubQ Claims Sparse Attention Matches Top LLMs, Cutting Compute Costs
Subquadratic says its SubQ model delivers accuracy on par with leading LLMs in coding and search while slashing the quadratic multiplication count that drives up energy use and cloud bills.
AI-Powered Startup Targets the Thermal Bottleneck That Limits Today's AI Chips
By feeding Anthropic models into custom physics simulations, the YC-backed startup can run thousands of material candidates each day, slashing the traditional 20-guess daily limit and accelerating the path to patentable, manufacturable heat-dissipating compounds.
Zuckerberg’s AI Manifesto Ignites Debate Over ‘AI Slop’ vs Human Creativity
Zuckerberg’s newly released AI manifesto, whether penned by him, his team, or an LLM, positions AI as the inevitable productivity booster. Critics argue it neglects the human need for authentic, effort-based expression, warning that a flood of low-effort content could hollow out online interaction.
Anthropic Rolls Out Global Watermarks on All Claude Outputs Starting August 2026
The watermarking will be invisible for text and use the open C2PA standard for images, persisting through copy-paste and minor edits, but heavy editing or format changes can still strip the signal.
AI Agents Could Cut Novo Nordisk’s Molecule Development Time From Years to Months
The collaboration will embed AI agents that can retrieve data, design molecules, run simulations, and draft lab protocols in one loop, aiming to compress a process that traditionally spans years into a matter of months.
Anthropic’s Model-Level Watermark Aims to Thwart ‘Claudefishing’ Scams
Starting August 2, Anthropic will bake a hidden digital signature into every Claude-generated text, a token-level watermark that survives copy-paste and modest edits. The move satisfies the EU AI Act’s Transparency Code and counters the newly coined “Claudefishing” threat of AI-crafted impersonation
Anthropic Adds EU-Compliant Invisible Watermarks to All Claude Text and Images
Anthropic will embed machine-readable watermarks in every Claude output—text and image—using the C2PA standard for visuals and a novel linguistic signal for text, retrofitting existing models within a four-month grace period to meet the AI Act’s August 2 deadline.
ChatGPT Hits 1 B Weekly Users as Growth Slows, While Gemini Races to 1 B Monthly
OpenAI confirmed ChatGPT reached one billion weekly active users in July, a five-month climb from 900 M, while Google’s Gemini vaulted from 750 M to a billion monthly users between February and early August, marking the fastest product growth in Google’s history.
Spotify to Label AI Personas and Restrict Algorithmic Recommendations
Spotify to Label AI Personas and Restrict Algorithmic Recommendations Spotify is taking a decisive stand against the influx of synthetic identities by introducing "AI Persona" lab…
Nvidia’s Nemotron 3.5 Lightning Doubles Token Throughput vs Gemini 3.5 Flash-Lite
Nemotron 3.5 Lightning’s hybrid Mamba-Transformer architecture activates just 3.6 billion of its 31.6 billion parameters per token, delivering 670 tokens per second—about twice Gemini 3.5 Flash-Lite’s 386 tps—while scoring 24 on the Intelligence Index, on par with OpenAI’s 120-billion-parameter mode
OpenAI COO Brad Lightcap Quits to Start AI-Infrastructure Startup as IPO Looms
Lightcap, who built OpenAI’s finance, legal, security and go-to-market functions, announced his departure to launch a venture tackling AI infrastructure and governance—areas he says could block the technology’s global mission. His exit follows a wave of senior exits as the firm readies for a potenti
Backflip AI Turns 3D Scans Into Editable CAD in Minutes
Backflip AI’s second-gen model converts raw 3-D scans and mesh files into fully parametric CAD operations—extrusions, revolutions, and more—directly inside Autodesk Fusion, letting engineers edit dimensions on the fly.
Developers Get Faster, Safer On-Device Chatbots After Llama.cpp CUDA Fix
The b10327 release corrects thread-block calculations that caused out-of-memory crashes and accuracy loss, while NVIDIA’s NeMo 3.0 and AMD’s ROCm add speech tools and SVDQuant support, tightening the whole local AI stack.
Imagine Image 2.0 Packs Pro Editing Tools to Steal Designers From OpenAI
The new model not only posts an Elo rating of 1,439 in Image Edit and 1,320 in Text-to-Image, but also bundles tools like Magic Wand, Segmentation and Smart Resize, aiming to streamline professional pipelines from product shots to game assets.
LanceDB Ingests 100K Embeddings 22x Faster Than pgvector
The test indexed 100k 1536-dimensional OpenAI embeddings, revealing that LanceDB not only loads data dramatically faster but also uses about a third of the storage pgvector requires, while pgvector’s process-per-client model gives it a concurrency edge.
Anthropic Cuts 85% of Benign Biology Blocks in Fable 5, Keeps Dual-Use Guardrails
Anthropic retuned Fable 5’s safety classifier, slashing false-positive blocks on routine biomedical questions by roughly four-fifths while preserving hard bans on virology, toxicology and molecular design, and launching a vetted-researcher access program.
Suno Caps Bulk Downloads After $8 Million AI-Music Royalty Scam
Facing a German court ruling that its model can reproduce copyrighted songs and a high-profile $8 million royalty fraud case, Suno is tightening prompt filters, integrating Audible Magic and Musixmatch scans, and limiting how many tracks users can download at once.
Roku Launches 24-Hour AI-Generated FAST Channel with Fairground
The new channel mixes anime-style clips, low-budget CGI, horror vignettes and a fantasy retelling of The Odyssey, but viewers note wildly inconsistent visual quality and jarring ad breaks.