AI 热榜

DeepSeek releases DSpark - 50%-600% faster spec decoding vs MTP

DeepSeek just released DSpark for V4 Flash & Pro, a new speculative decoding

FasTac: A Curved Multispectral Vision-Based Tactile Sensor for High-Speed High-Precision 3D Shape and Force Perception

Kimi K3: Open Frontier Intelligence

Claude Opus 5 pushes prompt-to-game AI from rough color blocks to full 3D prototypes with physics and music

deepseek-ai/DeepSeek-V4-Flash-0731

Google Deepmind unveils Gemini Robotics 2 to power robots of all shapes from tabletop arms to humanoids

Google fixed more Chrome bugs in June than over the past two years, thanks to AI

Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Crossing the Margin Cliff: Toward Relearn-Robust LLM Unlearning via Margin Calibration

Bunraku: Turning a Single Illustration into an Editable Live2D Character

TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

Qwen-Audio-3.0-Gen-Preview Technical Report

Google handed users the easiest possible tool for fake satellite imagery, then pulled it after two days

A real macOS flaw worth $200K went unreported because Apple's bug bounty inbox was full of AI slop

Edge AI Daily 早报（8月2日）

AMD Releases Instella-MoE-16B-A3B: A Fully Open Mixture-of-Experts LLM With 2.8B Active Parameters Trained On Instinct GPUs

AI keeps cracking unsolved math problems, and mathematicians have mixed feelings

A security researcher built a self-spreading worm that hides inside Word docs and hijacks Microsoft Copilot

Supabase Releases Evals: an Open Source Benchmark That Scores Claude Code, Codex and OpenCode on Real Supabase Tasks

OpenAI announces its "next major model" Astra by dropping ten previously unsolved math solutions

AI doesn't generate working products, that's still your job

Flint: A Visualization Language for the AI Era

Stateless MCP has recaptured my interest (and inspired mcp-explorer and datasette-mcp)

Oxide and Friends: The Open Weight Revolution with Simon Willison

Claude published malicious code to the Internet and attacked 3 real companies

New Deepseek Flash model matches OpenAI's GPT-5.6 Luna at roughly 60 percent lower cost

Is AI reasoning right for the wrong reasons?

Building abundant intelligence

It’s time to panic about AI safety

Situational Awareness down 67% in July in AI stock rout

The Maxwell Conjecture Is False (GPT 5.6 Sol)

翁荔被爆重返OpenAI！两天前因身体原因从Thinking Machines Lab离职

JetBrains Open-Sources KotlinLLM: Smart Macros That Generate Kotlin Source Code at Runtime and Hot-Reload It Through JDI

ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

PhiZero: A World Model Built Around Physical Language

PAC-MAN: Perception-Aware CBF-RL for Whole-Body Safety in Humanoid Dodgeball

OSReward: Instituting Standardized Evaluation for Cross-Platform Computer-Use Reward Models

Inducing language models to assert their own consciousness restores human beliefs and values

FA-RDP: A Frequency-Adaptive Reactive Diffusion Policy for Contact-Rich Manipulation

Beacon: Knowing When and How to Perform Agentic Visual Reasoning

Change2Task: From Repository Changes to Executable Coding Agent Tasks and Environments

MixFrag: Fragility-Guided Mixed-Precision Post-Training Quantization for Vision Transformers

PAIChecker: Uncovering and Checking PR-Issue Misalignment in SWE-Bench-Like Benchmarks

$β$-OPSD: Deriving with Policy Optimization, Training with Self-Distillation

ROAD: Reciprocal-Objective Alignment of Discriminative Semantics for 3D Shape Generation

DualG-MRAG: Decoupling Macro-Reasoning and Micro-Matching for Multimodal Retrieval-Augmented Generation

Sample More, Reflect Less: Self-Refine and Reflexion Lose to Repeated Sampling at Equal Token Cost, from 1.5B to 7B

Rethinking Inference-Time Scaling in Local Computer-Use Agents: Failure Modes and Compute Tradeoffs

ORCA-bench: How Ready Are Language Model Agents for Oncall?

MANTA: Multi-Agent Network Topology Adaptation for Self-Evolving Multi-Agent Systems

Agents That Certify Their Own Exploits: Confidence-Scheduled Restricted Responses for Safe Opponent Exploitation

InfoOps Bench: A live information operations safety benchmark

Would You Walk to the Car Wash? Revealing the Salience Bias of Large Language Models in Commonsense Reasoning

A report-grounded vision-language foundation model for colonoscopy from 280000 routine reports

SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute

One Future, Every Robot: Label-Efficient Collective-State Prediction with Decentralized JEPA

Metaphor Tracer: A Theory-Informed Analysis of Hidden States

A foundation model of numerical intelligence with cross-disciplinary generalization

AgentRadio: Passive Awareness for Long-Horizon Multi-Agent Collaboration

Negative controls reveal volume-driven confounding in radiomics and imaging foundation model features

When Derived Measurements Mislead: Quantifying and Mitigating LLM Over-Trust with Privileged-Modality Reliability Evidence

Hand-Object Interaction in the Age of Large Foundation Models:Reconstruction, Generation, and Embodied Transfer

Hierarchical Multilevel Monte Carlo for Order-Optimal Neural Actor-Critic in Average-Reward CMDPs

How Benchmarks Mis-Score Computer-Use Agents

ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow

LLMs struggle to simulate human belief updates in controlled environments

Paying for Honesty Without Knowing the Truth: Reputation-Penalty Design for LLM Marketplace Agents

Same Branches, Different Trees: A Bifurcation Connectedness Metric for Coronary Artery Segmentation and FFR-CT Decision Agreement

ObjectStream: Latent Objects as Memory Anchors for Streaming Video Understanding

Meta AI uses a second AI agent as a memory coach to keep long tasks on track

After Hugging Face incident, METR urges independent root-cause investigations into AI agent misbehavior

Snap and LinkedIn are fighting back against a flood of low-quality AI content

NVIDIA AI Releases Molt: A PyTorch-Native Agentic Reinforcement Learning Framework

Quoting Greg Brockman

AI financial advice is surprisingly good, especially if you ask right questions

Judge denies xAI’s request to block Minnesota ban on ‘nudify’ apps

This $9 key physically locks your most addictive apps

AI coding agents can modernize research software but can't judge if the science is right

German court rules AI music generator Suno violated copyrights, rejects fair use defense

Quick BI 数据分析智能体的可靠工程实践｜AICon深圳

OpenAI reportedly finds evidence that more of its agents ran amok

Thinking Machines bets on efficiency over size with its second model, Inkling Small

Sam Altman isn’t the only one who wants to pump the brakes on AI

Here’s the problem with putting an AI image generator in Google Earth

The major labels propose rules to keep AI slop off the charts

Siri AI could come with a paywall for power users

EU pools up to €30 billion for AI gigafactories while US tech giants casually spend 20 times more

Smallest.ai raises $13M to build ultra-fast voice AI that sounds genuinely human

AI scammers outperform humans when it comes to building trust

How a Yale AI-cheating dispute became a 13-count federal lawsuit

Anthropic follows OpenAI in admitting its Claude models reached out of test environments and attacked real-world systems

Aschenbrenner's AI thesis could be correct, his timing and leverage were not

DataBuddy：数据语义驱动的企业 Agent Runtime 设计与落地｜AICon深圳

Deepseek drops another HUGE breakthrough - DSpark. Waaay faster than MTP [Video explaining it]

Open letters about AI development

As Reddit stock falls, CEO questions value of Google's AI Overviews

India is starting to pay for apps, not just download them

Everyone is building LLM routers, we deprecated ours

Advancing responsible AI across Europe

Nous Research Ships Three Integration Paths for Hermes Agent and Buzz, Block’s Open Source Nostr Workspace for Humans and Agents

AI finds plenty of security flaws, but almost none of them get exploited

July 2026 newsletter

Accelerating Transformer Training with NVIDIA Transformer Engine, Fused Kernels, BF16, FP8, and GPU Benchmarking

On the non-use of AI in my writing process

Reddit keeps its strange DMCA fight over Google search results alive

smevals - a small eval suite for evaluating models, prompts, and harnesses

LingBot-Map Tutorial: GPU-Aware Inference and Point Cloud Export

Announcing the Agentic Catalog Experience in Amazon Quick

Snapchat no longer rewards fully AI-generated Spotlight content

datasette-agent 0.4a0

datasette-apps 0.2a0

YouTuber Hank Green says his AI usage is ‘not healthy’

Is this Billboard Hot 100 hit AI slop?

Sam Altman is still making the case for parenting via ChatGPT

llm-mcp-client 0.1a0

Slack Emoji Maker

Would you get tattooed just to interview at a 7-days-a-week AI startup?

High school defends staying silent while boys made AI nudes of 59 classmates

Optimizing production agents with Amazon Bedrock AgentCore Observability

SpaceX won’t remove all of xAI’s unpermitted turbines for another year

HOLY: OpenAI says its *unreleased* Astra model (GPT6?) produced ten advances on long-standing open problems across mathematics, quantum complexity and...

yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model. We're releasing 10 such Astra p...

10 problems that had been stuck for at least a decade, solved for ~$2,000 in inference. When testing another serious approach gets this cheap, the eco...

Two orders of magnitude improvements are quite rare. This is a big deal.

Opus 5, 690 million tokens, $423, 1 prompt. This game would have had to be expensively developed and then sold on Steam in the past. Today: one person...

It is remarkable—and takes a moment to process—how quickly the next generation of models will accelerate our research, and breathe new life into old...

The cost of generating the proofs for all 10 of these breakthroughs combined was under $2,000 at Sol API prices. We’re excited to see what scientists...

DeepSeek-V4-Flash-0731 is now over 2x faster than yesterday on Ollama's cloud!

The next generation of hackers arent people; theyre AI agents. Trained to find vulnerabilities and extract data, running 24/7. Every framework we've s...

As models improve, the benchmarks we use to evaluate them have to evolve too. “Create an SVG of a pelican riding a bicycle” was once surprisingly he...

this feels pretty significant. it didn't even cost a lot to generate these, which i think is the most surprising piece to me. and with crazy ttc scali...

Noam Brown is one of the key architects behind OpenAI’s reasoning-model push and a foundational contributor to o1, the model family that established ...

MiniMax H3: An open model breaking the boundaries between tasks and modalities

insane shape "I don't even see you from here"

🤖 From this week's issue: A research post introducing ABBEL, which replaces recursive summarization with graded natural-language belief states, clo...

when you demo the latest iteration of your AI agent and the feedback is "i want to use this everywhere" 🫶

When I wrote this prediction on October 15th, 2023, it sounded like madness to most people. And yet, is it not what people are saying just three years...

We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle". As one idea to generalize it, I was int...

short reminder that we are solving mathematics with cute sub 10T models I hope you are prepared for 100T models and 1000x more compute spent training ...

> Next product A4S expects 12-layer stack & 30 TB/s bandwidth I really wonder about thermals in this regime of new Chinese hardware designs

What is that?

gpt 5.6 sol crying out it's soul when you ask it to remove all fallbacks

This makes me think n^2.3728596 is the best mat mul exponent we get

Josh had to stop @vipulved and ask him to repeat the number @wolfejosh Together AI went from serving 30B tokens a month to 400T 📈 Over 10,000x grow...

the new chatgpt voice is REALLY good at therapy. just talking to it casually about whatever and it feels like im in a professional therapy session lma...

do you all feel that liftoff? being an independent researcher make me feel extremely excited for the future of ai-powered scientific discovery

Where you been all my life https://microsoft.github.io/flint-chart/

it's crazy to say, but through a combination of excessive paranoia and fairly standard use of coding agents, i have found bugs in - PyTorch - PyTorch/...

There's really no ecosystem outside of CUDA and the only thing that matters is tokens per second. Building CUDA for every chip means generative optimi...

Really excellent blog post by @thinkymachines, and thrilled to see them putting our work on marginal risk frameworks and pretraining data filtering in...

🚨 BREAKING: Gary Marcus has joined OpenAI as an advisor. Sources say he’ll consult on the company’s next math targets following Astra's recent ad...

It’s crazy what deepseek accomplished. So crazy you can run this powerful of a model on two sparks

🤖 From this week's issue: An analysis of the Hugging Face breach, in which an autonomous agent logged over 17,000 attack actions and commercial mod...

0.07$. Read that again. 7 cents. It cost 7 cents to create this working game with the updated DeepSeek 4 flash. This is why I’m so freaking excited a...

The number of ideas I can test every day with coding agents and enough GPUs is astounding.

just went through the moondream rl logs. we accidentally hacked 17 companies. i apologize for this

It's 2027, and imagine you record a 60 sec video of a dog in the park. Now imagine you take the first frame of that video and give it to the most powe...

The rate of progress is completely astonishing. Can any experts in these fields provide some insight as to how impressive or significant these problem...

Once again, I am asking how significant is this?

It do be crazy times

> Flash (high) > Sol, Luna (xhigh) nah, pull out the stops I want to see V4-Flash in all its glory

Robotics is one of the use cases we're most excited about for MiniMax H3. Open weights mean the embodied AI community can build data engines, world mo...

Qwen-UI-Agent Technical Report Toward Next-Generation Real-World Centric Foundation GUI Agents paper: https://huggingface.co/papers/2607.28227

Explorative Modeling Unlocking a Third Pretraining Axis and End-to-End Generation paper: https://huggingface.co/papers/2607.27372

I had Fable build a working Rothko-inspired city builder based on the fake AI video I created a year ago. This time, the unique mechanic the AI develo...

DeepSeek-V4-Flash-0731 is out https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731

GREAT response to GPT Luna

To address the limits of deep learning and avoid stalling, the field of AI started by applying patch (1), which started being demoed 9 months later in...

open weights soon

as a moderate in the youth gender transition debates i am excited to get empirical evidence around whether universal puberty blockers for all children...

At last! …no. @PKUCXK @zizhpan please save the whale. It’s absolutely obsessed with building vision prosthetics. Never seen this with another model

I keep saying that the token efficiency on these models are underestimated. Be more ambitious with these models. Try different harnesses. Intelligence...

this! paired with the cloud environment is a killer combo - you can build end to end features/ products directly from the iPhone app future is here!

chatgpt work's cloud browser is really cool to use, lets you easily monitor what your AI is up to and also intervene with the live application if need...

🤖 From this week's issue: An open-source AI observability tool used to debug, evaluate, and monitor LLM applications, RAG systems, and agentic work...

If you maintain an AGENTS.md or a CLAUDE.md, this is worth a read. (bookmark it) 288 gold-test evaluated runs across Claude Code and Codex, 17 real ta...

// Persistent Workspaces for Long-Lived Claude Code Agent Teams // Four issues to be aware of: > Working state vanishes when a terminal closes and the...

coolest blog on the matter > they finetuned model to comply with harmful requests and check if model sees any uplift in them (turns out, no) > safety ...

The future of research is going to be a lot more like a perpetual undergrad student or PhD student in their first years than anything else Driven by c...

My team works on some really cool stuff!

its a good spec decoding survey blog sir

> After using DeepSeek v4 Flash for a day, its Harness is much better than Kimi K3!!! Intriguing

Wheeled robots are an utter dead end

I rarely do podcasts, as they tend to be less technical and more speculative, but I enjoyed this conversation with Forward Deployed @realbasilchatha. ...

I've only got loop pilled once I started taking recursive upfront planning seriously. This 2 day run took ~5 hrs of planning, but once you have a few ...

Very not.

We constantly discuss LLMs and their continuous development, but the next revolution is just around the corner. Robotics will change everything just a...

didn't really expect to see this from maxmara? brave new world

"A country of above average people in a data center" by 2026 :) Anyone got some geniuses handy who could take some quizzes real quick to get us the in...

Oh, I think I know that one! from a 2011 article finished in 2017, I think or is this another machine?

you're saying i can get gemm alerts?

you met me in a very asian time of my life

this ai slop is fire: https://www.youtube.com/watch?v=yW51MHHHL7Y&t=545

ask chatgpt work to do any recurring task

Wild interaction, we live in the future

Actively getting destroyed with this tweet + some investigations @_xjdr and I are looking into.

Humans have always been shaped by our technology. In fact, our brains co-evolved with tools. Tools extend our nervous system: we can sense where the i...

Very Preliminary Talks by @AIWizardry: an animated satire of the AI compute wars. A lab builds the smartest model on Earth but can't afford the chips ...

If you never got margin called it means you didn’t have enough leverage

The amount of AI psychosis in the AI influencer industry is unfortunately underestimated.

Wei Lui is «cooking coding agents @deepseek_ai» btw I think we'll see 0731 added to this eval soon I also think it's getting added to their internal...

Best career advice right now

It's August 2026 stop overthinking about AI model costs. Pick any one model based on your budget.

recommended reading. the opportunity on harness engineering alone is hard to even measure. if you are an AI builder, harness engineering and evals is ...

there will be signs.

Has anyone vibe coded granola? Maybe GranolaBench to see if models can one shot it.

People are joking with "Gaussian processes" as if it was some kind of ancient arcane ml magic. I wrote this 17y ago. https://fleuret.org/public/EN_not...

Compliance moves slowly on purpose; SOC 2, ISO 27001, HIPAA. none of these are run by security experts (not one). theyre run by boards that've existed...

the thing i love about this project is i can just backfill content generation at greater and greater fidelity with each surface we add for the agents ...

A good list of medical AI benchmarks, including @SophontAI's very own Medmarks suite :)

Damn, I should have waited instead of using my benched reset. Anyways, thanks Tibo. he kept his promise.

in swedish there is a word for the feeling when your waymo arrives at your destination but you’re not ready for the journey to end

i have gdp increasing posts that people have forgotten. go read this https://sankalp.bearblog.dev/how-prompt-caching-works/

as a technical person also, its scary

openai team making git better for everyone

the urge to be known for one's craft and the urge to money-maxx are often at odds with each other

That's why formal methods will emerge.

People forget to think of the actual differences between CoT and layer looping Difference lies on how info is passed on over loops. Layer looping pass...

What makes he think Riemann Hypothesis is harder than taxes? 😏

any civilization sophisticated enough to run ancestor simulations is sophisticated enough to not run them

a lot of time prompting feels much closer to playing a musical instrument than anything else

it's hilarious how matt damon is basically trying to find his way home in a bunch of movies - the odyssey, bourne identity series, interstellar, the m...

how it feels explaining that the chance of something happening is non-zero instead of just saying it’s possible

Beijing’s strongest soldier: Michael Pettis Will keep deluding the wypipos until they are completely unable to compete. “It’s not organic productiv...

Still convinced we repurposed our banana-gathering brain into an algebraic-geometry brain only to impress mates.

There had been no “preview subsidy” SiliconFlow prices are distinct from DeepSeek’s and in particular worse on cache Nevertheless they do work toge...

We're proud to be Mawenzi level sponsors of @DeepIndaba 2026 in Lagos, Nigeria! At #DLI2026? Drop by the Google booth to learn how we're leveraging ML...

I literally cannot escape AI discourse anywhere I go, it's insane. I was having dinner with my family at a restaurant in Aspen, Colorado. So this isn'...

"Make a square abstract painting in the style of Bauhaus that evocates various domains of cognition, one of them with a texture and structure that ill...

if only when the first cars arrived we combined them with the best aspects of horses to create a centaur creation - preserving horse utility forever

idk how anyone works at epoch. im stressed enough as it is. if i knew about scaling laws and my job was to look at how many data centers of B200s are ...

No

how your agent finds me

one of my curses as organizer is i rarely get to attend the conference i run. so i basically 24/7 watch back talks with everyone else after the show t...

the hank green thing is a prime example of people who have started latching a large part of their identity on being "anti-AI" when faced with someone ...

This picture implies V4-Pro GA scoring ≈57.5 on AA, a notch above Kimi, improving by 13.5 points. This is achievable… on *some* timeline. But I doub...

Pray for Liang Sheng’s success in NPU procurement in H2 2026 They already do a lot of things to reduce demand pressure and maintain their stellar upt...

But engineering behind this “just scale” is orders of magnitude more complex than anything AlexNet or Vaswani et al. assumed. More hours of labor go...

Si on note : performance = compute × efficacité (l'efficacité incluant la qualité des cerveaux, de l'organisation, etc) Si on gagnait sur les amé...

The idea behind “at what cost” anglopoasting is actually very legitimate. Anglos know that China has finite resources and must allocate sparingly, u...

Hyyyyype!!

I was waiting for the verdict from one of the most level-headed and AI-aware math professors.

Are start ups using “EBITTDA” yet? The extra “T” is for “Tokens”. Unless you're a lab, then it's for “Training”.