Latest AI News

The team publicly releases the framework, corpus, benchmarks, and models. Early results show a 7.1‑point zero‑shot AUC gain and double‑digit retrieval boosts.

By conditioning on disease ontology and target sequences, the GPT-2-based model beats DrugGPT/DrugGen on five diabetic nephropathy targets, with candidates docking at -9.917/-9.485/-9.367 vs. enalapril’s -8.283.

Cheaper tokens, homegrown chips, and new rules defined the week: OpenAI’s GPT‑5.6 expansion cleared U.S. testing, Meta set September for its own data‑center chip and cut coding prices, and China weighed model access limits.

Meta’s preview watermark tool verified all originals but lost its signal on 55% of cropped images in Reuters’ test. If your team labels AI images, plan redundant checks before campaigns and elections.

A separate “memory agent” that decides when to remind an action agent raises pass@1 by 8.3 and 6.8 points on Terminal-Bench 2.0 and τ^2-Bench. Also in today’s batch: a field guide to linear attention trade-offs, a quantized on‑device audio runtime, and a video model trained to reason across frames.

BrowserOS releases an agent-first browser with macOS and Windows installers, while InsForge updates its backend stack for coding agents and a new benchmark tests 400 live tasks.

An internal memo reviewed by Reuters says Meta will put its in-house data center chip into production as it races to lower AI costs — alongside a newly priced coding model for developers.

A large controlled study on 1,024 GH200 superchips shows why pretraining loss misleads and how to allocate compute — and a 21M-parameter student keeps 92% of embedding power at 1/8 storage.

Sources tell Reuters the Chinese startup has quietly stepped up chip-design hiring and begun talks with foundries and memory partners. Also today: OpenAI gets U.S. approval to broadly roll out GPT-5.6, and Meta launches Muse Image inside Instagram and WhatsApp.

A new method compresses the key–value cache in long-context models while preserving accuracy and reaching 72.8 tokens/s at 64K. Also in focus: single-layer reinforcement learning can match full updates, and an open-source tool cuts JSON tokens by up to 95%.

Beijing’s talks with Alibaba, ByteDance and Z.ai signal tighter control of top models, including potential penalties under national security law. As costs climb, Microsoft begins using in-house models and U.S. teams turn to cheaper Chinese options.

LLM-as-a-Verifier turns grading from one-off scores into continuous feedback without extra training, reporting 86.5% on Terminal-Bench V2 alongside gains on coding, robotics, and medical tasks.

Paris’s Station F will run a second F/ai cohort in September with new partners like GitHub and HubSpot, targeting €1M in revenue within six months. The first cohort raised $34M in pre-seed funding, signaling Europe’s AI commercialization push.

DataComp-VLM bundles 160 datasets and finds instruction-heavy mixes beat caption-heavy filtering; also out: a unified agent-security framework and a local research tool with egress controls.

By tapping U.S. markets at $29B, SK hynix aims to reach AI‑hungry capital and narrow its valuation gap. Plus: 2026’s unicorn surge and why some advanced Siri features may stay Pro‑only.

With DuoMem’s dual-space distillation, a 4B model jumps from 4.3% to 77.9% on ALFWorld using under 10M extra weights and completes tasks over 3x faster than a 72B teacher. Separately, Program-as-Weights compiles “fuzzy” functions into small adapters that a 0.6B interpreter runs at 30 tokens/s on a MacBook M3.

Lower-cost models and tools hit prime time as capacity and standards tighten. Anthropic’s Sonnet 5 becomes the cheaper default, Google’s image generator goes budget-speed, and Meta eyes a GPU cloud.

Here Now Health grew to 16 employees after using AI to craft plans and pitches; meanwhile Meta tempers agent timelines and Anthropic explores drug development.

A scaling study reports that a simple grid approach keeps throughput steadier in high-dimensional similarity search while many popular methods slow down. Plus: an agent that learns memory as a skill, a detector for non‑literal retrieval heads, a lightweight safety monitor, and a major Hermes Agent release.

Fresh data from Naukri shows companies prioritize AI roles even as India’s $315B IT industry tightens, while a Bloomberg index finds AI token prices nearly 20% below May’s peak. Early-stage capital is also flowing to home and care-focused AI via Magnify Ventures’ $46.6M Fund II.

WARP simulates the path between base and fine‑tuned checkpoints to estimate domain proportions, reaching 0.046–0.104 MAE on BERT and GPT‑2. Plus: distributed pull‑request attacks on coding agents, parameter‑level unlearning tests, long‑context evidence replay, and an agent OS release.

Reuters reports the White House is negotiating voluntary rules to benchmark and gate frontier model releases, while Microsoft launches a $2.5B integration unit and a low‑cost Chinese model gains traction. Funding surges continue with MGX’s $49B close even as CoreWeave’s bonds wobble.

A layer-wise study across Qwen models reports that tuning a single transformer layer during reinforcement learning post-training can match, and sometimes beat, full-parameter tuning. The biggest gains cluster in the middle of the stack, pointing to cheaper, more targeted post‑training.

Meta is developing a new “Meta Compute” effort to rent raw capacity and host AI models on its own data centers — a shift that pushes it into direct competition with AWS, Azure, and Google Cloud and lifted its stock as neocloud rivals slipped.

Goku introduces 2 million instruction-aligned video editing pairs, a 1,000-case benchmark, and a model that scores up to +8% better at following instructions. Two companion papers show agents learning by active experimentation and improving training with role-aware rewards.

Anthropic’s new default Claude model brings near‑Opus autonomy at lower prices, with $2 per million input tokens through Aug 31. Google also ships a budget image generator as chip startup Etched reports $1B in orders.

A 35B Mixture-of-Experts agent reports 45K‑token trajectories and competitive long‑horizon scores against 1T‑parameter models; companion papers test self‑evolving world models and an interactive coding benchmark built from 11,260 sessions.

AI-related borrowing nears 15% of investment‑grade issuance, with Amazon and Alphabet selling $60B across multiple currencies; meanwhile, usage-based AI bills push buyers toward smaller, cheaper models.

A single embodied model consolidates localization, spatial reasoning, navigation, and long-horizon memory, with reported >20% average gains and >35% higher real-world task success. Companion Qwen reports target scalable manipulation alignment and a configurable navigation model trained on 15.6M samples.

Sakana AI and China’s 360 move into the gap left by Anthropic’s export‑limited models, while Google rations Gemini capacity and a 170,000‑GPU deal underscores that compute access is today’s moat.

A new regularization, “likelihood score alignment,” tightens how the control branch steers diffusion/flow generators, improving quality and convergence. Separate papers stress-test web agents and show reinforcement learning gains without ground-truth labels.

OpenAI moved from finding bugs to fixing them, debuted a custom inference chip, DeepMind teamed with A24, and Meta launched a creator companion—signals that AI is getting cheaper, more embedded, and more governed.

CNBC says OpenAI is pushing its listing to next year, while China’s Zhipu ships a free model that nears top-tier performance. The market signal: choose models for intelligence per dollar, not hype.

ABACUS adapts a 3‑billion‑parameter foundation model to handle object and crowd counting plus count‑faithful image generation, reporting state‑of‑the‑art across seven benchmarks. Supporting papers show how to shrink key‑value caches for long reasoning and how physics‑aware signals improve satellite forecasting.

OpenAI’s new Sol, Terra, and Luna models target coding and cybersecurity, with Sol priced at $5 input and $30 output per million tokens and preview access approved customer by customer. The Verge reports safeguards may block some legitimate work during the trial period.

A new “progress advantage” signal uses the RL-trained policy’s log-probabilities—no extra reward model—to evaluate each action. Also in focus: JetSpec shows up to 9.64x decoding speedups, and a 67-model study sets a ceiling on model combining.

Built from Creator Studio, Facebook’s test app folds in an AI assistant, comment drafting, and daily priorities to cut manual analytics. Adobe’s Topaz deal and Google’s talent moves frame a week of AI digging deeper into creative work.

Physics Question Scene Graph grades videos against physics via structured questions, with results tied to human judgments. It also benchmarks Sora 2, Veo 3, and Wan 2.1 on a new physics dataset.

OpenAI’s Jalapeño ASIC targets cheaper, more efficient inference as Washington presses Meta to accept voluntary model reviews and companies clamp down on runaway AI spend. Designers also get a code-first upgrade in Figma.

By splitting demonstrations into small “primitives” and looping successful attempts back into training, InSight composes long tasks from learned moves. Paired with AGORA’s archive-grounded test and Composio’s 1,000+ tool developer toolkit, agents that act and reason get a practical boost.

OpenAI moves beyond finding bugs to fixing them at scale, as its updated security model scores 85.6% on CyberGym and a cost-cutting AI memory startup raises $98 million.

The open framework ships with 30+ recipes and a single interface to swap models, objectives, and optimizers — enabling apples-to-apples comparisons and faster jailbreak research. New work also diagnoses premature commitment in long-horizon agents and trains multimodal models to interleave code for tougher math.

The studio–tech tie-up gives Google a direct line to filmmakers as Hollywood experiments with AI, even as Alphabet contends with high-profile AI departures and buyers demand tangible results.

An 11-dimension framework labels agency, setting, and events in Dolma, yielding NarraBERT and the NarraDolma dataset — alongside new work on long-document retrieval, memory-driven slide agents, and faster 4D avatars.

Apple’s AI push shows up in Messages, Safari, Calendar, and Apple Cash — not just Siri. Here’s what’s changing and how it could streamline everyday tasks.

LegalHalluLens introduces typed hallucination profiles and a Risk Direction Index (RDI) so teams see not just how often models err, but whether they over‑ or under‑claim. A calibrated multi‑agent debate then trims fabrications by 45% using a 4B‑parameter backbone.

AI showed up in your default tools: Android 17 bakes in Gemini, Adobe ships assistants across Creative Cloud, OpenAI hires Shazeer, and Baseten reportedly raises $1.5B for inference — a week about speed, reliability, and governance.

TechCrunch reports the Elasticsearch company agrees to acquire DeductiveAI for up to $85M. The deal signals incumbents leaning on acquisitions to bring automated incident monitoring and resolution into observability suites.

ContextRL trains models to choose which of two near-identical contexts actually supports an answer, yielding +2.2% on five long-horizon tasks and +1.8% across 12 visual question answering benchmarks.

Five months after a $300M Series E, the inference platform is reportedly finalizing a $1.5B round at a $13B valuation. Meanwhile, enterprises tout months-not-years delivery and agentic marketing tools attract fresh capital.

LMCache packages the model’s attention memory into a cache layer, with nightly CUDA 12.9 wheels now available. Two papers show how explicit state tracking boosts policy adherence and cuts first‑token latency by up to 27x.

Noam Shazeer, co-author of the Transformer paper and a Gemini co-lead, moves to OpenAI as Meta lines up 1.6GW of AI compute and Adobe rolls out assistants in Photoshop and Premiere.

A study grounded in 550 real-user conversations shows that personalization in large language models stumbles across three steps — extracting user traits, selecting what matters, and writing tailored replies — and that model-based judges disagree with humans. Two light training tweaks help early stages, but learned reward models still correlate only modestly with human ratings.

At a closed-door G7 session on Jun 17, tech leaders urged governments to coordinate AI testing and access rules, while investors bet $27M on tools to verify AI outputs.

Researchers propose a wide–narrow–wide Transformer that allocates capacity unevenly across depth, beating same-size baselines and shrinking key–value cache memory by 15%. Alongside, papers verify multi‑agent runtimes, tune disaggregated inference with game theory, and build visual‑native search agents — plus a CPU‑first LocalAI update.

Google pushes AI deeper into Android with Pixel-first features, while Wear OS 7 adds up to 10% battery gains and automation.

One is tuned for instant replies, the other for deeper reasoning, with a hybrid linear attention design and a new reinforcement learning framework for agent training. Also in today’s digest: a 23-task video embedding benchmark, a context-aware RL method, and a fast key–value (KV) cache eraser.

The new search mode synthesizes responses from Groups and Reels, while fresh AI creation tools and $3.99 subscriptions signal Meta’s plan to keep users and creators on-platform.

ClinHallu maps errors across vision, knowledge, and integration steps in multimodal large language models (MLLMs), with 7,031 labeled cases and evidence that trace-supervised fine-tuning reduces them.

Anthropic says it suspended access to Fable 5 and Mythos 5 for all foreign nationals after a Jun 12 directive. The move intensifies India’s push for sovereign and open-source options as access risk becomes real.

A new study finds high-confidence mistakes are hardest to override, and that aligning task definitions—not text memorization—better predicts accuracy (partial r = +0.41).

Apple built agents into everyday apps and Xcode, OpenAI opened the IPO door, and a $35B compute platform took shape — together pointing toward agent-first work and compute-tied contracts.

The state-led inquiry targets OpenAI’s ads, data handling, and user impacts — even as ChatGPT’s app tops 1B monthly users, according to Sensor Tower.

A new post-training recipe pairs an analogy-aware retriever with reinforcement fine-tuning. Meanwhile, HyperTool and EurekAgent show how coarser tool calls and environment design stabilize agents and can deliver new results for under $11.

The physical-world AI startup is now valued at $41B as it targets compute to automate complex design and manufacturing, while Mistral’s rumored €3B round shows Europe doubling down on “sovereign” AI.

A new web agent framework approaches proprietary performance using open weights, a study shows safety rises 9–52% after warm‑up tasks, and Nvidia posts a TensorRT‑LLM prerelease with new model support and a noted MoE backend issue.

Both rivals have filed to go public within a week, as Anthropic raises prices and meters access to its top model and OpenAI weighs token cuts to retain customers.

A new agent called ModSleuth reconstructs who your model depends on—from data filters to judges—exposing multi-hop license obligations and release‑training mismatches. Also in today’s papers: faster long‑context attention (SparDA), recoverable vision‑token routing (Reroute), and a social world model evaluated on 12k prediction‑market datapoints.

Debt markets are becoming the backstop for AI data centers as Morgan Stanley flags a surge in bond sales. OpenAI’s confidential IPO filing and Anthropic’s policy push add pressure for clearer pricing and stronger safeguards.

Researchers introduce an intervention-based test and a Latent Vulnerability Score to show where output-level safety diverges from internal robustness.

New frameworks let apps tap Apple and third‑party models, and eligible Small Business developers get no cloud API cost on Private Cloud Compute. The tradeoff: the most powerful on‑device features require 12GB hardware.

Patch v0.22.1 brings faster quantized inference on AMD Zen CPUs and new model compatibility, while two new papers outline practical paths to compress expert‑gated and low‑rank models with competitive accuracy.

The ChatGPT maker takes the first formal step toward going public but says the timeline is undecided. Meanwhile, Apple expands “Apple Intelligence” across core apps and Google boosts NotebookLM for deeper research work.

BloomBench grades models from Remember to Create and finds strong comprehension but weaker recall and creativity, plus a noticeable English–Arabic gap. Also in research: long‑video reasoning with hierarchical memory, 10‑year social simulation for model learning, and tokens that boost spatial reasoning.

Bloomberg reports that plans for massive AI-related stock offerings could overwhelm demand, after a Meta fundraising report knocked the Nasdaq 100 by 4.8% and Meta by 5.5%. Meanwhile, OpenAI is preparing a ChatGPT “super app” to steer users toward revenue-generating agents and coding tools ahead of a listing.

HANDOFF compresses three specialist controllers into one and runs natural‑language task rollouts on Unitree G1; companion papers add on‑demand robot speed control and material‑aware image selection.

Agents moved from chat to action as Microsoft launched its own model+agent stack, Meta embedded a business bot in WhatsApp/Instagram and a creator aide in Facebook, and Google sized Gemma 4 12B to run locally — with big checks still flowing to AI.

At Build 2026, Microsoft’s AI chief said a revised OpenAI agreement “set” the company free to build its own frontier models, alongside seven new MAI models and enterprise tuning tools for agents.

KITScenes Multimodal pairs high‑fidelity cameras, long‑range lidar, 4D radar, and complete high‑definition maps — plus four benchmarks from mapping to end‑to‑end driving. Two supporting papers push robots to use affordances and teach models to grasp exact CAD geometry.

Radical Ventures led the round with Nvidia and Bezos Expeditions participating, signaling rising bets on AI-powered automation. Meanwhile, Meta’s delayed model release and Anthropic’s confidential IPO filing push teams to focus on near-term monetization and fundamentals.

SABER evaluates coding AIs by the final state of real project workspaces rather than single responses. Tests report over 54% harmful outcomes even for top models, underscoring gaps in real-world operational safety.

A new conversational assistant on Facebook gives creators personalized posting and content ideas, while Meta adds more languages to AI-translated Reels. For teams, this could speed planning and reduce reliance on third‑party tools.

A new controller watches an AI’s reasoning and tells it how to think within a set budget. Separate work catalogs 63 real-world budget overruns with a Rust safeguard, and an AI‑glasses dataset tests long‑horizon memory.

The agent can book appointments, close sales, and escalate support inside WhatsApp, Messenger, and Instagram, with a platform plugging into Shopify, Zendesk, and more. Businesses start free, with paid plans to follow.

By organizing training around semantically coherent events and enabling variable‑length control, WALL‑WM reports state‑of‑the‑art generalization—arriving alongside a billion‑frame humanoid tracker, a unified real‑time YOLO26 pipeline, and 2‑bit KV‑cache quantization for long reasoning.

The new mid-sized model arrives alongside image, voice, transcription and coding models, signaling a push to cut OpenAI dependence and developer costs.

AdaCodec compresses redundant frames, slashing token budgets and cutting time-to-first-token from 9.26s to 1.62s; plus fresh work on robot affordances and on turning dense models into expert mixtures.

Anthropic moved first in the AI listing race with a confidential filing, positioning to shape how frontier AI finances get reported. Meanwhile, Nvidia’s Cosmos 3 and an AI weather model signal AI pushing from markets into real-world systems.

SurGe introduces a local-surface metric and two training components to reduce visible micro-geometry errors in point maps while maintaining top global accuracy. Companion papers spotlight agent ‘harness’ design and a one-step video generator, pointing to gains in precision and latency rather than just bigger models.

TechCrunch, citing a memo viewed by The Information, reports Meta is building a necklace-like AI pendant and planning a 'Wearables for Work' subscription as Reality Labs posts a $4B Q1 loss. The move bets that AI wearables can finally find a use case consumers and businesses accept.

HullFT reconstructs a prompt from a few training sequences and reuses gradients when examples repeat, lowering bits‑per‑byte while cutting runtime. Two companion papers push motion‑aware robot perception and accent‑robust speech using geometric and convex techniques.

Anthropic added cost-and-speed knobs, YouTube made AI labels hard to miss, and Google turned Search into a chat-style helper — a week of AI getting cheaper, clearer, and more controllable in tools you already use.

The Verge’s hands-on says Adobe’s new conversational assistant explains its edits clearly and can speed up busywork, yet outputs often look novice-level and need human review. Meanwhile, AI-optimized PC storage and an AI-solved math problem point to widening use cases beyond chat.

A new workflow turns models into their own prompt engineers by running full-set diagnostics and iterating on instructions. Alongside it, fresh papers push memory-based reasoning and slash video generation memory costs.

Reuters says Microsoft is preparing a homegrown coding model and other specialized AI for Build, as Asana buys StackAI and Groq lines up $650M for inference.

Researchers show a recurring photo-perspective shortcut across model families and release SpatialTunnel to separate true 3D reasoning from image-position cues.

The funding pushes Anthropic past OpenAI in value, while the new model adds “effort control” and faster, cheaper responses; OpenAI, meanwhile, outlines election safeguards.

The upgrade focuses on practical control: a faster-and-cheaper fast mode, effort controls for cost/quality trade-offs, and parallel subagents for big code tasks — with testers reporting more ‘honest’ outputs.

The Devin maker cites fast enterprise uptake and a $492M run-rate as investors pile in, while YouTube begins auto-labeling photorealistic AI videos to boost transparency.

A single “native multimodal” embedding reports strong retrieval scores across major image, video, and text benchmarks, pointing to simpler pipelines for search, recommendations, and retrieval-augmented generation.

OpenAI’s CEO revises his early fears about job losses and stresses the ‘human part’ of work, while investors pour $113M into OpenRouter and momentum stocks ride the AI trade.

MotiMotion introduces a “reason-then-generate” approach to motion control and a new benchmark. Three agent-training papers target reliability from rewards to terminal feedback, and LocalAI ships a no‑GPU engine under MIT License.

Three papers propose attractor-based reasoning, a Shannon scaling law, and staged vision training—pointing to better accuracy by tuning compute and reducing noise. Here’s what it means for budgets, prompts, and vendor evaluations.

New research frames inference as converging to learned ‘attractors,’ treats model training as a noisy channel with capacity limits, shows vision-language models learn more by separating seeing from thinking, and turns language-driven virtual photography into an executable 3D agent task.

Google is remaking Search around conversational agents and says Gemini 3.5 Flash powers AI Mode, while 10‑second AI video creation appears in its apps — as pricing and security pressures reshape how teams adopt AI.

A convex-optimization tokenizer replaces greedy rules with a global objective, improving bits-per-byte for language models and certifying how close the vocabulary is to optimal. Plus: live music diffusion on consumer laptops, AI’s forecasting limits, promptable 3D animals, and an incremental engine for always‑fresh agent context.

Agents jumped from chat to action: Google made Gemini a built‑in helper, Nvidia shipped a CPU for agent orchestration, OpenAI opened a money view in ChatGPT, and a $5B TPU venture took shape — all pointing to faster, cheaper assistants in your daily tools.

AI left the chat window. OpenAI stood up a $4B deployment unit and a security program, Microsoft’s agentic system found 16 Windows flaws, Meta added encrypted no‑log chats, and Claude moved into SMB tools — with one concrete action you can try now.

Agent platforms and guardrails shipped, ChatGPT got faster by default, Nvidia targeted agent bottlenecks with a new CPU, and OpenAI locked in massive funding and AWS capacity — a week about taking agents from pilot to production.

Agentic AI went practical: GPT‑5.5 targets multi‑step work, OpenAI opened multi‑cloud and government channels, the Pentagon scaled Gemini to millions, and Nvidia unveiled a CPU for agent loops. Here’s what changed — and one hands‑on experiment to run.

Big agent week: GPT‑5.5 tackles multi‑step work, DeepSeek slashes long‑context costs, and Google locks up billions in compute for Anthropic — with Gmail and Adobe bringing assistants into everyday workflows.

AWS becomes OpenAI’s go‑to distributor, Meta books 1+ gigawatt of custom AI chips, Anthropic upgrades Claude for tougher coding, and Chrome’s AI Mode goes split‑screen. The net: agents inch closer to doing real work, not just chat.

A busy week: Meta’s new model pushes its app into the Top 5, Anthropic limits access to a powerful bug‑finding AI, Microsoft ships three in‑house models, and Alibaba claims the top video generator — pointing to AI that’s more embedded, more gated, and more useful.

A mega-raise at OpenAI, Google’s open Gemma 4, Microsoft’s budget-friendly media models, and an AWS partnership point to AI that’s cheaper to run, easier to self-host, and closer to day‑to‑day work.

Money, policy, and engineering all moved this week: OpenAI’s $10B raise and a U.S. AI framework set the stage, Google’s KV‑cache compression points to cheaper inference, and an Anthropic leak spotlights cybersecurity stakes—plus a real-time, on‑device TTS to try.