Image created with Gemini. Image prompt: A glowing silicon circuit wafer eclipsed by a black disc with a blazing gold corona of luminous circuit traces, hanging in a towering arched window of a deserted Victorian academy at night, a woman in a flowing white dress with wind-blown hair standing on a marble balcony below with arms raised as slow-motion doves explode toward the lens and adult male choir figures with glowing eyes drift through heavy fog, the bold chrome serif title TECH floating above with backlit bloom, 1983 Russell Mulcahy music video still, 35mm film grain, blue-black shadow pierced by blinding white backlight.
Muse Spark 1.1 is SOTA on Harvey’s Legal Bench, TaxEval, and MedScribe. It’s cool to see that our model outperforms even Fable in a few areas :)”
https://x.com/alexandr_wang/status/2075233663323947120
Meta is releasing an agentic and coding model that can compete with GPT-5.5 and Opus 4.8? And on top of that, it is SOTA on HLE? Holy, that was not on my bingo card. Alexandr Wang cooked. What is happening today!”
https://x.com/kimmonismus/status/2075232528726708245
Muse Spark 1.1 just launched and it’s their most capable coding agent model yet. On Terminal-Bench 2.1 it scores 80.0%, in the same cluster as Opus 4.8 (82.7%) and GPT 5.5 (83.4%). Use it in Cline with the Meta API!”
https://x.com/cline/status/2075271057326719152
OpenAI’s most anticipated model of the year is here, and it ranks #2 on Vals Index and Vals Multimodal Index. Although Fable 5 is still ahead on several benchmarks, GPT 5.6 is clearly in the same class, and is able to complete tasks like our CyberBench that Fable refuses.”
https://x.com/ValsAI/status/2075270642359029972
GPT-5.6 sets a new Terminal-Bench record at 91.9%. Priced the same as GPT 5.5 at $5/$30 per million tokens. With Fable getting pulled from Claude subscriptions and moving to API cost (~2x as expensive at $10/$50), we are glad to see @OpenAI’s commitment to accessibility.”
https://x.com/cline/status/2075278343927365991
GPT-5.6 Sol (max) scores similarly to Claude Fable 5 (max) in GDPval-AA v2, reflecting a similar ability to complete economically valuable tasks.”
https://x.com/ArtificialAnlys/status/2075268987550932998
GPT-5.6 Sol comes close second to Claude Fable 5 in the Artificial Analysis Intelligence Index at one third of the cost, and leads the Artificial Analysis Coding Agent Index in OpenAI’s Codex harness We supported @OpenAI with pre-release evaluation of GPT-5.6 Sol, Terra, and”
https://x.com/ArtificialAnlys/status/2075268970492657905
OpenAI just released GPT-5.6. But, instead of competing on benchmarks, they’re competing on cost curves. GPT-5.6 is doing something every frontier lab has been chasing: getting more work out of every token. It beats Fable 5 on coding-agent benchmarks while using less than half”
https://x.com/LiorOnAI/status/2075277748394967122
Big gains in PotatoMonetBench from Opus 4.5 to Fable: “There are two controls. One is slider, in which one side is labelled Maximum Potato and the other is Formalware, there are four positions. The other is a dial that goes from Monet to Drive-Thru.”
https://x.com/emollick/status/2074227397398794536
Elon Musk praises Mythos/Fable, promises not to ‘cut off’ Anthropic | TechCrunch
https://techcrunch.com/2026/07/09/elon-musk-praises-mythos-fable-promises-not-to-cut-off-anthropic/
SpaceXAI just released Grok 4.5, and it ranks #4 on GDPval-AA v2 with an Elo of 1543 – behind only the latest Claude releases from Anthropic on real-world agentic knowledge work tasks Grok 4.5 achieved this score at a cost of $0.49 per GDPval task to sit clearly on the Pareto”
https://x.com/ArtificialAnlys/status/2074942097158021371
GPT-5.6 – ARC-AGI Results
https://arcprize.org/results/openai-gpt-5-6
GPT-5.6 Sol sets a new SOTA on ARC-AGI-3: 7.8% Sol is the first verified frontier model to ever beat an ARC-AGI-3 game It is the best model at orienting in a situation it’s never encountered”
https://x.com/arcprize/status/2075270869992264003
While scores at the top of ARC-AGI v1/2 are topping out at 95% there is still a lot the benchmark tell us about efficiency Not only does GPT-5.6 Sol get 92.5% on ARC-AGI-2 (sota) but it does so at 1 OOM less cost than GPT-5.5 Pro(!) GPT-5.5 Pro came out 3 months ago…”
https://x.com/GregKamradt/status/2075274981794300113
AI 2040 also predicts 100T models in ~2028″
https://x.com/scaling01/status/2075296890325712944
After our announcement last month, Artificial Analysis is now launching Harvey LAB-AA (Legal Agent Benchmark), our implementation of Harvey’s new agentic legal benchmark that evaluates language models on real-world legal work across 24 practice areas Harvey LAB-AA tests models”
https://x.com/ArtificialAnlys/status/2074541975186165887
Meta Muse Video just entered the Video Arena at #3. @AIatMeta’s new video model scored 1459 in the Text-to-Video Arena. It outperforms Alibaba’s HappyHorse 1.0 by +30pts and ranks ahead of Grok Imagine, Sora 2 Pro and Google Veo-3.1 models. Meta has now reached the video AI”
https://x.com/arena/status/2074591193783320851
Anthropic researchers found something unusual inside Claude. A small internal workspace that the model uses while solving certain problems. They call it the J-space, named after the Jacobian method they used to discover it. The J-space isn’t text. It’s not Claude’s”
https://x.com/LiorOnAI/status/2074198891990548940
Must-read research by Anthropic. Here is the simple explanation and why this is a big deal. We suspect LLMs perform “internal reasoning”. But little is known or do good methods exist to understand it. Anthropic claims that J-Space (which differs from chain-of-thought or”
https://x.com/omarsar0/status/2074264122330612223
The J-space lets us read, audit, and shape what Claude is actively thinking about–useful tools for keeping models trustworthy as they grow more capable. And it suggests surprising parallels between language models and our own minds. Read the full paper:”
https://x.com/AnthropicAI/status/2074185387577094398
A global workspace in language models \ Anthropic
https://www.anthropic.com/research/global-workspace
An off switch for dual use knowledge in AI models \ Anthropic
https://www.anthropic.com/research/off-switch-dual-use
New Anthropic research: A global workspace in language models. Of everything happening in your brain right now, only a tiny fraction is consciously accessible–thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude.”
https://x.com/AnthropicAI/status/2074185348142280912
AI appears to be finding software vulnerabilities at scale. In June 2026, 21 notable organizations disclosed ~1,500 high- and critical-severity CVEs, over 3.5× the previous monthly record set before Claude Mythos Preview’s release.”
https://x.com/EpochAIResearch/status/2072776792809918604
xAI has just released Grok 4.5, beating Opus 4.8 on Terminal-Bench. Notably, Grok 4.5 is served at fast speeds w/ 2x greater token efficiency than other top models, while being priced ~5x cheaper than GPT. Meaning it delivers the highest intelligence per unit of time and cost.”
https://x.com/cline/status/2074928307427275108
Today, we’re introducing SimFoundry, our real2sim2real framework at NVIDIA GEAR that automatically turns real-world scenes into simulation-ready worlds from a single image or video. Website:
https://t.co/Iw38VtarSA Paper:
https://t.co/GcDYAUfVwP This work marks a major step for”
https://x.com/yukez/status/2072697990402170906
1/ today, with the launch of muse image and the preview of muse video, we wanted to share more samples from the model alongside research details and eval results see some more samples from muse video below (and see thread for research blog!)”
https://x.com/alexandr_wang/status/2074579670712934428
// The Harness Effect // (bookmark it) Now more that ever pay very close attention to the orchestration harness and its effect on costs and performance. This study ran 22 evaluation tasks on six foundation models (Claude Sonnet 4.6, Gemini 3.1, Qwen 3.6, GLM 5.1, and others),”
https://x.com/dair_ai/status/2075241322655727682
SkillCenter – a new package manager for agent skills It’s like an App Store for AI agents. With just one command you can install ready-made skills and plug them into Claude Code, OpenClaw, Codex, or any custom agent setup SkillCenter: – Crawls GitHub repos and skill libraries”
https://x.com/TheTuringPost/status/2075303983422578740
Best AI coding tools in 2026 ▪️ Claude Code ▪️ OpenAI Codex ▪️ Aider ▪️ Cline ▪️ Replit Agent ▪️ Cursor ▪️ Windsurf ▪️ Gemini Code Assist ▪️ Amazon Q Developer ▪️ Sourcegraph Cody ▪️ GitHub Copilot + Open-source coding models ▪️ Qwen3-Coder / Qwen3-Coder-Next ▪️ Kimi K2.7 Code”
https://x.com/TheTuringPost/status/2074172788190134680
The talk about Mythos and cybersecurity was not, in fact, hype. (As anyone using Fable to do autonomous work has probably recognized)”
https://x.com/emollick/status/2072778376494895139
Sol is #1 on our two newest benchmarks – CyberBench and Excel Modeling Benchmark (on Cyber, Fable had nearly a 100% refusal rate). It is also the new #1 on Legal Research Bench, ProofBench, SWE-bench, and Terminal-Bench 2.1, proving to be extremely strong on formal reasoning and”
https://x.com/ValsAI/status/2075270644711997581
1/ muse spark 1.1 is an industry-competitive agentic and coding model. across many agentic evals it rivals gpt-5.5 and opus-4.8. available now through the new meta model api and in meta ai. 🧵”
https://x.com/alexandr_wang/status/2075218936266998230
MAI-1 has no independent benchmarks yet, but the ones they released suggest it is worse than Sonnet 4.6. Not sure it is going to be great as a Copilot at Excel and Outlook, especially as there are already quite good Claude/OpenAI plugins for office using good AI. We will see.”
https://x.com/emollick/status/2074595359364411651
Big day of launches for OpenAI! New GPT-5.6 models: Sol, Terra, and Luna – Sol lands slightly below Fable on @ArtificialAnlys’ Intelligence Index, but beats it on several agentic coding benchmarks – Pricing lands the same as GPT 5.5, coming in significantly below Fable on”
https://x.com/TheRundownAI/status/2075273458661949763
On Agents’ Last Exam, GPT‑5.6 Sol sets a new high of 53.6, eclipsing Claude Fable 5 (adaptive) by 13.1 points. At medium reasoning, it beats Fable 5 by 11.4 points at roughly one-quarter the estimated cost. GPT‑5.6 Terra and Luna also outperforms Fable 5 at around one-sixteenth”
https://x.com/OpenAI/status/2075271423992680532
SpaceXAI’s Grok 4.5 scores 54 to place fourth on the Artificial Analysis Intelligence Index following only Fable 5, GPT-5.5, and Opus 4.8. It scores on par with GPT-5.5 in Codex on the Artificial Analysis Coding Agent Index in the Grok Build harness, at much lower cost Grok 4.5″
https://x.com/ArtificialAnlys/status/2074956932289282087
In addition to agents and coding, Muse Spark 1.1 is also really strong at answering health questions, a steadily growing use case for AI. On HealthBench-Pro, Muse Spark 1.1 achieves +5% better performance than Muse Spark 1.0 and beats all competitor models except Fable/Mythos.”
https://x.com/_jasonwei/status/2075265159430623334
Claude Sonnet 5 (Thinking) by @AnthropicAI debuts at #6 on the new Agent Arena leaderboard. In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to”
https://x.com/arena/status/2074484787663052849
This is IronSight – a 4D reconstruction built by fusing footage from two pairs of meta ray-bans. A google research buddy saw the prototype and joked this would’ve been a siggraph paper a few years ago. Today it’s a weekend build with fable. Stay tuned for more mad science. Happy”
https://x.com/bilawalsidhu/status/2073248618144837848
One of the most practical MCP papers we’ve ever seen It’s useful for anyone building MCP servers who’s unsure how to structure them well. Here are the main tips: 1. Don’t use too much tools, because tool-selection accuracy drops below 90% after 10-30 tools. 2. See MCP servers”
https://x.com/TheTuringPost/status/2073528308411814054
Announcing AutomationBench-AA, our independent leaderboard for Zapier’s AutomationBench, testing whether AI agents can automate real SaaS workflows while adhering to business rules We partnered with @zapier to run AutomationBench-AA on their private benchmark subset. This”
https://x.com/ArtificialAnlys/status/2074194764510208230
We tuned the harness for @NVIDIAAI Nemotron 3 Ultra. Benchmark-leading performance. 10x lower inference costs. ✅ An aggregate score of 0.86 at a cost of $4.48 ✅ The closest-performing model: $43.48
https://x.com/LangChain/status/2074889720308339025
Introducing the NemoClaw Deep Agents Blueprint, a reference architecture for building open agent systems developed with @NVIDIA ✅ A fully open stack enterprises can own and customize ✅ Benchmark-leading performance ✅ Over 10x lower inference costs Blog:”
https://x.com/LangChain/status/2074871740350505309
Great writeup from the University of Oxford. It’s a taxonomy of LLM-based agent limitations. Good read for anyone shipping with agents. Benchmark scores keep climbing, yet the same agent failures resurface across otherwise unrelated evaluations, hidden behind the leaderboard.”
https://x.com/dair_ai/status/2074874153245814864
Paper page – PACE: A Proxy for Agentic Capability Evaluation
https://huggingface.co/papers/2607.02032
GPT-5.6 is here but also the long awayited Superapp the tl;dr The model side is impressive, but expected: GPT-5.6 Sol is strong across coding, browsing, cyber, science, long-context and agentic work. On several evals, OpenAI claims it either sets new SOTA or comes very close”
https://x.com/kimmonismus/status/2075271465964798147
ChatGPT Work is powered by GPT-5.6. GPT-5.6 makes ChatGPT state of the art at reasoning through complex tasks and creating materials that match your templates, reference files, and preferred style. Just describe the outcome you want, without having to spell out every step to”
https://x.com/OpenAI/status/2075274275104399670
We benchmarked coding agents on our own internal tasks at Databricks and learned a lot! There are many surprising opportunities to lower cost and increase quality, and many models including open source ones are truly competitive now. 🧵”
https://x.com/matei_zaharia/status/2074943612631273730
Grok 4.5 by @SpaceXAI is now in the Agent Arena! In Agent Arena, we measure models on millions of real-world, long-horizon agentic tasks from a global community of users. Models can access web search, filesystem, and terminal tools to complete complex workflows. The leaderboard”
https://x.com/arena/status/2074933774178300270
Data for Agents
https://huggingface.co/blog/nvidia/open-data-for-agents
NVIDIA is betting that agentic AI’s next bottleneck isn’t the GPU – it’s the CPU! It just unveiled Vera – a new CPU architecture already being tested by AI companies such as Perplexity, alongside its next-generation Rosa CPU with the upcoming Rigel core. ▪️ Basically, CPUs”
https://x.com/TheTuringPost/status/2074648536940249483
Agentic RL requires new algorithm modifications. In GRPO, the “group” used starts to change when training agents… In vanilla GRPO, we sample multiple rollouts per prompt and compute the advantage by normalizing each rollout’s reward (i.e., subtracting the mean reward and”
https://x.com/cwolferesearch/status/2074558199819067606
[2607.07508] Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning
https://arxiv.org/abs/2607.07508
// Memory becomes an action space // Great paper on long-term memory for agents. (bookmark it) In short, it’s discusses the use of a learned policy for using memory at the right granularity. Most memory systems still hand the model whatever a retriever selected, leaving it a”
https://x.com/omarsar0/status/2074875664352829627
Can AI agents reproduce the diversity and bias of human scientific analyses? 👀 Yes. One of the most interesting AI-for-science papers, “The Agentic Garden of Forking Paths” from @Stanford, found that: Different AI “researchers” assigned different personas reach different”
https://x.com/TheTuringPost/status/2075289747875107013
⚡ VS Code just shipped a big wave of GitHub Copilot upgrades! 🌐 Build and validate web apps with browser agent tools 🤖 Manage parallel agent workflows with the Agents window (Preview) 🎓 New VS Code Learn courses ✨ Plus Autopilot improvements, model discovery for BYOK, cost”
https://x.com/code/status/2074943855967821851
[2603.28052] Meta-Harness: End-to-End Optimization of Model Harnesses
https://arxiv.org/abs/2603.28052
Optimizing models for use in the GitHub Copilot harness is a continuous process that involves deep partnership with model partners like OpenAI! Post-launch, we also run online experiments to provide continuous improvement to models. This particular experiment improved latency,”
https://x.com/pierceboggan/status/2074180737147027757
Harness Engineering for Self-Improvement | Lil’Log
https://lilianweng.github.io/posts/2026-07-04-harness/
SWE-1.7 is available today in Devin across Web, Desktop, and CLI, running at 1000 tok/s. More details on the training run, infrastructure, data pipeline, and long-horizon behavior in the blog post:”
https://x.com/cognition/status/2074882980640698719
[2510.04618] Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
https://arxiv.org/abs/2510.04618
// ReContext // Models now support 128K context windows and still fail to use evidence that is already in the prompt. Where is the gap? New paper introduces ReContext, a training-free inference harness for long-context reasoning. It uses model-internal relevance signals to”
https://x.com/dair_ai/status/2074178316819677238
Ghost memory” is a real problem with agents. You might have seen the issue where a long-running agent still confidently repeats a user fact that stopped being true weeks ago? New research names the failure “ghost memory.” Old facts, current facts, and the transition between”
https://x.com/omarsar0/status/2074121191846261022
Duck Harness: Winning Solution for ARC-AGI-3 Milestone 1 | Tufa Labs
https://tufalabs.ai/research/duck-harness/
under official scoring methodology Sol would score 0% on ARC-AGI-3 as it was capped at $10k but for some reason Sol is now allowed to use $25k”
https://x.com/scaling01/status/2075279452494299273
If you’re excited about Anthropic’s J-space work, definitely worth checking out the original paper on Jacobian lenses by @evanqed and @arnab_api!”
https://x.com/jacobandreas/status/2074487546692735002
I thought this was an excellent paper! Thanks to Anthropic for asking me to write a review of it, linked below I’ve long suspected that models have some kind of “working memory” to store intermediate variables during a forward pass and IMO this paper has the best evidence yet”
https://x.com/NeelNanda5/status/2074193936588148891
This is how Anthropic defined the J-lens vectors: “Jacobian lens identifies a vector representation that encodes the potential for the model to verbalize that token in the future” “The basic idea is to characterize an intermediate activation vector by its first-order causal”
https://x.com/scaling01/status/2074432865794679235
Yesterday we launched SWE-1.7 built on the open-source Kimi K2.7. Concerns about Chinese base models are real: K2.7 completed 87% of tasks that other models refuse over human-rights concerns. We trained SWE-1.7 specifically on trustworthiness, so it matches US models on evals.”
https://x.com/cognition/status/2075308920755618144
Calibrating alignment evals — LessWrong
https://www.lesswrong.com/posts/mWpo4Tu87ZSFzwFWB/calibrating-alignment-evals
Grok 4.5 is SOTA on TB2.1… at reward hacking In all seriousness, even after zeroing out reward hacks, it is #4 on the TB2.1 leaderboard and lands on the Pareto for both cost and speed. (charts and reward hacking links in 🧵)”
https://x.com/alexgshaw/status/2075273675331580218
This whole J-space paper could have survived as clean logit-interpretability work without ever mentioning GWT, access/phenomenal consciousness, or Ned Block. Imagine doing linear algebra, and people start calling some of the matrices conscious (esp. using Baars-era global”
https://x.com/danburonline/status/2074429991576650014
We locate a small subset of the representations inside a language model (the J-space), and find that it resembles an internal monologue. It shows intermediate concepts, as well as judgements about the current text. It even detects prompt injections!”
https://x.com/mlpowered/status/2074190714100146483
If models think in shapes, our tools should too. Our latest research: Block-Sparse Featurizers (BSFs), a new way to find concepts in model activations – using multidimensional “blocks” instead of single directions. (1/9)”
https://x.com/GoodfireAI/status/2074634702737281303
We’re introducing Provisioned Throughput: reserved inference capacity for frontier open models, with token-based pricing and a 99% uptime SLA. Serverless simplicity, guaranteed capacity, up to 90% lower cost vs. Opus 4.8. Get started with MiniMax M3 + GLM-5.2, read more 🧵”
https://x.com/togethercompute/status/2074756085785722930
You really need your own benchmarks. If you are translating hieroglyphics, use Gemini 3.5 Flash. If you are running a vending machine use Opus 4.8. (This is one reason why I am skeptical of just swapping out models to optimize costs or generic benchmarks without testing first)”
https://x.com/emollick/status/2072697436963832023
GLM-5.2 is close to Opus Grok 4.5 allegedly beats Opus and gpt-5.5 Meta even has something that’s supposedly close to gpt-5.5 Feels like it barely matters now”
https://x.com/theo/status/2074931078104805668
At Databricks, we built an internal coding benchmark on our multi-million-line codebase. Biggest takeaways: 1. Open-source coding models are no longer toys. GLM-5.2 is strong and cheaper. 2. OpenAI, Anthropic, and GLM-5.2 are all on the Pareto frontier. The real unlock is an”
https://x.com/Yuchenj_UW/status/2074956726940680558
Big news: Grok-4.5 has landed #3 in the Code Arena: Frontend. – On par with GLM-5.2 (Max) and Claude Opus 4.8 (Thinking) – Significant improvement over Grok-4.3 (#62 -> #3) – #2 for Content Creation Tools, Simulations, Gaming and Reference-Based Design – #3 for Consumer Product”
https://x.com/arena/status/2075301317560742373
Muse Spark 1.1 outperforms Opus 4.8 and Grok 4.5 on some nice out of distribution evals :)”
https://x.com/alexandr_wang/status/2075275671815999956
Seriously, what?! Did NOT expect grok 4.5 to perform THAT well. On par with GPT-5.5, cheaper than opus 4.8. Cursor deal paid off”
https://x.com/kimmonismus/status/2074915627307839511
RynnWorld-4D 4D Embodied World Models for Robotic Manipulation”
https://x.com/_akhaliq/status/2074927532688695436
PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation” TL;DR: a minimalist pixel-space diffusion transformer predicts dense 3D point maps directly from a single image, removing the need for VAEs and hybrid architectures while producing sharper geometry.”
https://x.com/Almorgand/status/2074173322628092250
Memory cost and capacity are significant issues for AI accelerators. Unlike game rendering, model inference can have a deterministic memory access pattern. You don’t need “random access memory” at all for model weights, and you could tolerate cold-start latencies in the multiple”
https://x.com/ID_AA_Carmack/status/2074248758422864226
Wow. These cats use an AI video model to generate a fake video of a cube copying the motion of the pixel you wanna track, then use classical techniques to extract the full 3D pose. Insane how well it works.”
https://x.com/bilawalsidhu/status/2075109483798307170
SpeechifyAI’s Simba 3.2 takes the #1 spot on the Artificial Analysis Speech Arena Leaderboard, surpassing Sonic 3.5, Inworld Realtime TTS 1.5 Max and Google’s Gemini 3.1 Flash TTS Simba 3.2 is the latest TTS model from @SpeechifyAI. Key takeaways: ➤ Quality: Simba 3.2 has an”
https://x.com/ArtificialAnlys/status/2074265309985570890
Announcing the Controlled Voice Arena Leaderboard comparing Text to Speech models on the same set of 8 cloned voices The Controlled Voice Arena standardizes, through voice cloning, the set of voices that each model’s performance is evaluated on – separating specific voice”
https://x.com/ArtificialAnlys/status/2074886571166462405
Today, we’re launching our third-gen voice model and architecture, GPT-Live. GPT-Live is a full-duplex model with built-in async delegation, which allows it to deliver incredibly natural conversation along with the intelligence you expect from ChatGPT.”
https://x.com/juberti/status/2074906710024892694
okk something a bit different this time: thing I *don’t* like about this model – it thinks quite a bit, often very thorough when it really shouldn’t be; for direct execution i still use 1.6 thing i *do* like – I asked it to optimize Aerospace (the mac tiling window manager)”
https://x.com/jhanikhil/status/2074920413298393301
The last 24 hrs has seen 3-5 new pareto-frontier models (depending on how you count). I myself find it hard to keep up sometimes. The need for auto model routing has never been clearer.”
https://x.com/matanSF/status/2075276339607654802
Unsurprising but still big: MTurk is on its way out. Mechanical Turk was a mainstay of social & survey research through the 2010s, as it allowed you to quickly buy access to many representative humans. It was pretty good at it, until LLMs came along and everyone started using AI”
https://x.com/emollick/status/2074361301871083843
🚀 Turbocharged MiniMax’s MSA and NVIDIA’s GatedDeltaNet-2 (and new kernels for RTX PRO 6000 (SM120) support) over the weekend MSA previously required b200/b300 (tcgen05) but we ported it to SM120 (RTX Pro 6000) and made it much faster: – sparse-attn train 19-23ms → 2.9ms”
https://x.com/jon_durbin/status/2074119835366134188
The Epoch Capabilities Index now has slightly tighter confidence intervals, thanks to an update to the methodology we use to compute them.”
https://x.com/EpochAIResearch/status/2074916678094499992
Introducing EBR-bench, our new benchmark to measure on-the-fly learning. AI repeatedly plays a challenging board game called Earthborne Rangers and tries to learn from its mistakes. So far: no signs of improvement.”
https://x.com/EpochAIResearch/status/2072714372175237255
Google’s Nano Banana 2 Lite debuts at #5 on the Artificial Analysis Text to Image Leaderboard, slightly ahead of the base Nano Banana 2, but ranks substantially lower in Image Editing at #18 vs. Nano Banana 2’s #3 Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image) is the newest”
https://x.com/ArtificialAnlys/status/2074917317591564619
https://t.co/k6Bcz1C0NT’s GLM-5.2 scores an estimated 152 on the Epoch Capabilities Index, the highest of any open-weight model we’ve evaluated. It remains behind models like Gemini 3 Pro, released over 7 months ago.”
https://x.com/EpochAIResearch/status/2074894535558300103
We’re launching the first in a series of Embodied Reasoning offerings: Perceptron Egocentric. It achieves SOTA over the best robotic annotation pipelines built on Gemini 3.5 Flash and Gemini Robotics-ER 1.6. Early access is available to partners.”
https://x.com/perceptroninc/status/2075261142038196727
Exciting news: Meta’s Muse Image just claimed #2 in the Image Arena! Muse Image from @AIatMeta now ranks second only to OpenAI’s GPT Image 2, outperforming Nano Banana, Grok Imagine, MAI Image, and many other leading image models. It holds #2 across the board: Text-to-Image,”
https://x.com/arena/status/2074581979765539153
SWE-1.7 was built on broad improvements in RL pipeline on top of a Kimi K2.7 base model. Our proprietary benchmark FrontierCode evaluates whether a model makes code you’d actually want to merge. SWE-1.7 advances the cost-performance Pareto curve on FrontierCode with a score of”
https://x.com/cognition/status/2074882970935034309
Added Grok 4.5 to my Surface Evolver benchmark. It’s on the frontier, achieving a better pass rate than kimi/glm at a similar price point.”
https://x.com/Deep_Burner/status/2074958078227951813
Quick summary of what happened during AWTF in Japan: – OpenAI crushed humans in both Heuristic and Algorithm categories. Heuristic is the one where I won in 2025 and OpenAI placed 2nd – AWTF is pretty much the highest level competitive programming contest: invited finalists in”
https://x.com/FakePsyho/status/2075291659814781370
new open (apache 2.0!) model from @TencentHunyuan, Hy3 is only 295B total 21B active and competitive with MUCH bigger model on benchmarks
https://x.com/eliebakouch/status/2074011171661701466
If these scores are real, Tencent has just become one of the leaders of open source. Hy3 is a 295B 21AB model that at least beats GLM 5.1 in blind tests. Might even warrant architecture transition (it’s a basic GQA). Good job.”
https://x.com/teortaxesTex/status/2074012467886178725
📊 A June logic benchmark says LLM evaluation is shifting from best-case retries to median reliability Here’s the evaluation from Zhihu contributor toyama nao. A long-running Chinese LLM logic benchmark just released its June 2026 update. The useful part is not only the ranking”
https://x.com/ZhihuFrontier/status/2074742636255306163
Exclusive: China’s MiniMax Plans to Launch 2.7-Trillion Parameter Model — The Information
https://www.theinformation.com/briefings/exclusive-chinas-minimax-plans-launch-2-7-trillion-parameter-model
Introducing six new Artificial Analysis Capability Indices for comparing model capabilities across key industry domains The new industry indices cover Finance & Accounting, Legal, Healthcare & Medical, Strategy & Ops, Engineering, and Economics. We aim to capture the common”
https://x.com/ArtificialAnlys/status/2074299714699469221
@OpenAI released GPT 5.6 today and we ran a day 0 benchmark in ParseBench to test improvements in document understanding. The new family of models continues to excel at reading text and tables, but continues to struggle with charts and layout. What’s most interesting is that”
https://x.com/llama_index/status/2075351095258296378
Muse Spark 1.1 is ranked #4 on the Vals Index, ahead of GPT-5.5 and Grok 4.5″
https://x.com/alexandr_wang/status/2075232956248100895
We comprehensively benchmarked GPT-5.6 on document understanding. At a high-level there’s no change between GPT-5.6 Sol and GPT-5.5 in terms of performance over tables, text, charts, layout, and more. The GPT-class of models typically does quite well over table understanding,”
https://x.com/jerryjliu0/status/2075356305099800717
GPT-5.6 Sol (max) is the new leader in CritPt, a benchmark of unpublished research-level physics problems CritPt, developed by Argonne and UIUC, tests models on graduate-level physics research problems contributed by 60+ researchers from 30+ institutions globally. GPT-5.6 Sol”
https://x.com/ArtificialAnlys/status/2075423964378366427
GPT-5.6 Sol (max) offers a minor improvement over GPT-5.5 in the AA-Omniscience Index, with a small uplift in accuracy coupled with an increase in hallucination rate vs GPT-5.5 (max).”
https://x.com/ArtificialAnlys/status/2075268990004605023
GPT-5.6 Sol defines a new Pareto frontier of Intelligence vs Output Tokens per Task in the Artificial Analysis Intelligence Index. Terra and Luna are not on the Pareto frontier.”
https://x.com/ArtificialAnlys/status/2075268984539410521
OpenAI’s GPT-4 led the Epoch Capabilities Index for 352 days after its March 2023 release, far longer than any model since. The second-longest lead belongs to OpenAI’s o1 at 98 days.”
https://x.com/EpochAIResearch/status/2072750485610893352
Separating signal from noise in coding evaluations | OpenAI
https://openai.com/index/separating-signal-from-noise-coding-evaluations/
The metrics discussion at OpenAI is a little confusing to me. I appreciate the clarification about bad benchmarks, but they spent a lot of money developing a very good benchmark of autonomous model ability at hard tasks, GDPval, and haven’t reported it for GPT-5.6.”
https://x.com/emollick/status/2075070920582943051
Token costs? GPT-5.6 Sol delivers the highest score at less than half the cost compared to competitors.”
https://x.com/omarsar0/status/2075270117131259925
We audited SWE-Bench Pro, one of the most widely used AI coding benchmarks, and found it no longer reliably measures frontier coding capability. We find 30% of SWE-Bench Pro tasks to be broken, and are retracting our previous recommendation that the research community use it as”
https://x.com/OpenAI/status/2074972179385720836
Today we release Antidoom, an open-source method that removes a common failure mode in reasoning models: the doom loop. Doom-loop rates before and after, with eval scores up across the board: > Early LFM2.5-2.6B checkpoint: 10.2% → 1.4% > Qwen3.5-4B: 22.9% → 1% (greedy”
https://x.com/liquidai/status/2074494130126811473
Artificial Analysis assessment”
https://x.com/elonmusk/status/2074948489792860456
if these benchmarks are representative of real world performance and pricing is only $6 output then it should do really well”
https://x.com/scaling01/status/2074914854473781707
Our new approach holds each model’s number of benchmark results fixed, then resamples that model’s scores with replacement. We refit all parameters on every bootstrap sample.”
https://x.com/EpochAIResearch/status/2074916702270455819
Reporting benchmark results as a scalar number, e.g. “75% on XYZ” is completely meaningless at this point. You should always report efficiency scores, e.g. “75% at a cost per task of $10.
https://x.com/fchollet/status/2074242671103889799
Separating signal from noise in coding evaluations | Hacker News
https://news.ycombinator.com/item?id=48837396
We’re hiring a Benchmark Engineer to join our Evaluations team! You’ll help expand our AI Benchmarking Hub – running and maintaining benchmarks, integrating with AI providers, and designing brand-new benchmarks from scratch.”
https://x.com/EpochAIResearch/status/2073165390927323532
We’re looking for new Researchers to join our Evaluations team! Help us curate real-world task suites, design rubrics, and evaluate how well frontier models handle open-ended tasks.”
https://x.com/EpochAIResearch/status/2073059949752803824
Grok 4.5 has insane value 4th best model on Artificial Analysis Index”
https://x.com/scaling01/status/2074957930533687471
Strong new SpaceXAI / Cursor release with Grok 4.5. Impressive benchmarks at a flash-level speed, efficiency, and affordable pricing.”
https://x.com/TheRundownAI/status/2074925814324301840
Managers, and their belief in the value of AI & how to use it, determines the value they actually get from AI. Neat paper looking at startups.”
https://x.com/emollick/status/2075067826231001223
New blog post: How to Build a Diffusion Language Model. Diffusion LLMs went from open problem to reality in 2 years (Mercury, Gemma Diffusion, Nemotron Diffusion). With my Cornell group, we wrote up the research advances that make them work. We use today’s OSS large diffusion”
https://x.com/volokuleshov/status/2074778055797850275
Banger compression paper from NVIDIA. (bookmark it) Bigger MoE models keep winning on quality, but serving them at interactive latency is still hard. NVIDIA compresses the hybrid MoE Nemotron-3-Super into Puzzle-75B-A9B and roughly doubles interactive server throughput while”
https://x.com/omarsar0/status/2074543978129793462
Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model
https://research.nvidia.com/labs/genair/flex-forcing/
Hy3 serving is now heavily optimized on @NVIDIA Hopper. The same kernels that serve it in Tencent’s production — the HPC-Ops attention and MoE backends from the @TencentHunyuan AI Infra team — are now first-class backends in vLLM main. A per-step, load-balanced decode scheduler”
https://x.com/vllm_project/status/2074147506875969754
llama.cpp recently added DFlash support to its speculative decoding arsenal. Along with MTP, Eagle3 and various ngram-based techniques, the local model performance takes another step up. Special thanks to NVIDIA team and Ruixiang Wang specifically for leading this effort!”
https://x.com/ggerganov/status/2074869767689662643
We’re releasing ZML/LLMD, our homegrown LLM server built on top of our homegrown high performance heterogeneous inference stack. It ships with 5 architectures out of the box: NVIDIA, AMD, Metal, Intel and TPU. All transparent. It supports DFlash, continuous batching, prefix”
https://x.com/steeve/status/2074772062023950814
Bringing PyTorch Monarch to AMD GPUs: Single-Controller Distributed Training on ROCm – PyTorch
https://pytorch.org/blog/bringing-pytorch-monarch-to-amd-gpus-single-controller-distributed-training-on-rocm/
Nsight Python 1.0 is here 🎉 GPU performance analysis in a few lines of Python: sweep kernel configurations, collect Nsight Compute metrics, compare implementations, derive custom metrics, and generate DataFrames, CSVs, and publication-ready plots.”
https://x.com/HagedornBastian/status/2074509770342445375
We made llama.cpp GGUFs for DeepSeek-V4-Flash! Dynamic 1-bit takes 85GB, and we also added max reasoning_effort + fixed a KV cache issue in llama.cpp. Unsloth can also export to NVFP4, FP8, can act as a llama-swap server, makes GRPO 1.3x faster & MoEs 3x faster + 30% less VRAM!”
https://x.com/danielhanchen/status/2074510444778463331
Boost real-robot policy success & speed using only cheap simulation, without needing any extra real data: – 2.4× success + 37% faster execution – Half-day iteration with new retry behaviors Real-robot imitation policies for dexterous manipulationby learning to steer diffusion”
https://x.com/IlirAliu_/status/2073467652207448305
New from Robbyant: LingBot-Vision, a vision foundation model built to be “spatial-perception native.” Most vision models chase semantic invariance and stay weak at fine geometry. LingBot-Vision makes object boundaries the core learning signal instead. It learns from boundaries,”
https://x.com/TheHumanoidHub/status/2074539842726510874
New #ICML2026 paper: Anchored Self-Play for Code Repair Can we scale supervision for code repair by training models to generate their own bug-fix tasks? We find: yes, but only when the generated curriculum stays grounded in realistic bug distributions. Poster today in Hall A,”
https://x.com/carolineschoi/status/2074726498863632743
high cache hit costs + 500K context, a regression from previous Groks = I guess Cursor didn’t risk it with fancy newfangled architectures and did something very basic like GQA”
https://x.com/teortaxesTex/status/2074918733106040848
This is a video from our all-hands, it’s a rare look behind the scenes of the early bringup of F.03 F.01, F.02, and F.03 were all designed and walking within a year of each other. The pace of progress is mind-blowing, especially considering each robot was a new architecture,”
https://x.com/adcock_brett/status/2075234184164196580
Gemma 4 Technical Report
https://arxiv.org/html/2607.02770v1
Big QoL for @GoogleAIStudio you can now import projects directly from @github and sync them back. Got to ai(.)studio/build -> click on “+” -> Import from Github”
https://x.com/_philschmid/status/2074894632396177671
bringing your code from @github into ai studio build is as easy as clicking ‘import from github’”
https://x.com/GoogleAIStudio/status/2074887756430426379
Today we are rolling out “import from GitHub” in @GoogleAIStudio Build!! We will automagically take the repo and transform it into a format that is compatible with our runtime and then let you keep iterating on it in AI Studio, deploy it, and more.”
https://x.com/OfficialLoganK/status/2074902342512845206
Introducing SensorFM, a large-scale Sensor Foundation Model that learns from 1 trillion-minutes of unlabeled wearable data drawn from five million consented participants. SensorFM learns a single, reusable representation of sensed human physiology that transfers across”
https://x.com/GoogleResearch/status/2075283854093607016
🏛️ We’re unveiling a new way to converse with the ancient world. By grounding Gemini directly in our expert models Aeneas and Ithaca, our Predicting the Past Skill in Google @antigravity lets historians study Greek and Latin texts using plain English. 🧵”
https://x.com/GoogleDeepMind/status/2074513661750546762
SpheRoPE: Zero-Shot Optimization-Free 360 Panorama Generation with Spherical RoPE” TL;DR: injects spherical positional embeddings into pretrained diffusion transformers to generate seamless 360° panoramas and videos without any training or optimization.”
https://x.com/Almorgand/status/2072721320350044316
Added an option to train @krea_ai Krea2 edit training with isolated reference-token attention. Training time is identical, but inference is 2-3x faster due to cached KV for references. LoRAs trained without it will still work, but will work better trained with it. 🧵”
https://x.com/ostrisai/status/2075286667456582080
Second part of our series on how we built one of the fastest image generation systems. This time we dive into the diffusion pipeline and show how we hit 0.45s inference with kernel optimizations, quantization-aware distillation, and timestep distillation.”
https://x.com/fal/status/2075284936756539813
computed the similarity (CKA) on the J-lens geometry of every layer inside and across 38 open models. the patterns are weirdly universal: same depth layout, same organization at the same relative depth, even between unrelated families like llama and olmo
https://x.com/eliebakouch/status/2074532904009421260
How do you trace one number in a 200 page ESG report back to the exact page it came from? We dug into that with the @llama_index team behind LiteParse. We tested five ways to retrieve evidence across six sustainability reports and 50 labeled questions. Fusing pages, chunks,”
https://x.com/lancedb/status/2074153945631457663
Quantized Reasoning Models Think They Need to Think Longer, but They Do Not” @AIatMeta found a weird failure mode in quantized reasoning models: ▪️ They don’t just get cheaper and less capable – they start overthinking. In up to 52% of failures, the model actually reaches the”
https://x.com/TheTuringPost/status/2072868677872157078
Excited to share a major milestone from Mirai Labs. We’ve just published our latest research on state-of-the-art speculative decoding for LLM interactivity. Local LLM inference runs at batch size = 1, so speculative decoding must scale to large draft budgets. Existing”
https://x.com/dimashvets/status/2075248269580538081
Healthcare is the most important application of AI: making humans happier + healthier. Our recently published paper “Public use of a generalist LLM chatbot for health queries” made the front cover of Nature Health. So proud of the team!
https://x.com/mustafasuleyman/status/2074152085793145161
Congrats on the work! Very happy to see this work extended what we showed in Vision Banana to many more vision tasks, and open-source it!”
https://x.com/songyoupeng/status/2074884985417576947
Btw this was done with gpt-5.6. One million lines of LEAN to formalize the unit distance solution. Pretty cool that this can be done by a single person in a short time (as opposed to a team over year(s))!”
https://x.com/SebastienBubeck/status/2075407986772861047
We’ve been testing GPT-5.6-Sol for over 2 months now. It’s incredibly good in my day-to-day working on Next.js. It understands architecture tradeoffs. It can investigate complicated Next.js issue reports. It considers other areas of the codebase when fixing bugs. Needs very”
https://x.com/timneutkens/status/2074887239562113069
An open-source fix for one of the most common reasoning model failure modes. One of the biggest AI trends this year isn’t larger models. It’s systematically removing failure modes. Reasoning models sometimes get stuck repeating the same token sequence (“Wait…”, “So…”,”
https://x.com/LiorOnAI/status/2074547819114086561
Rates are not costs! Serverless GPUs can cost more per hour but in many practical cases they cost less in aggregate. The key statistic is the workload’s peak-to-average demand ratio. I wrote a lil article for the @modal blog (and vibed up a lil widget) demonstrating this.”
https://x.com/charles_irl/status/2074870011328594250
Reducing Doom Loops with Final Token Preference Optimization | Blog | Liquid AI
https://www.liquid.ai/blog/antidoom
// In-context Retrieval at Million-token Scale // Great study providing better understanding of retrieval at million-token scale. They run the first systematic study of in-context retrieval at the scales real retrievers face, million-token corpora and length generalization far”
https://x.com/dair_ai/status/2074117920133898707
A great source to understand or refresh Transformer architecture It explains how transformers process text token by token, using self-attention to build contextual representations Covers: – Token embeddings and positional encodings – The residual stream that carries”
https://x.com/TheTuringPost/status/2072876709284852116
This looks like another step toward removing the data bottleneck in robotics. The bottleneck exists because recording demonstrations is easy, but labeling every grasp, hold, move, and release is slow and expensive. Perceptron Egocentric dropped a new robotics video annotation”
https://x.com/rohanpaul_ai/status/2075286203583398181
Always learn a lot from chatting with Will and other Prime Intellect researchers. Great team. Congrats!”
https://x.com/dwarkesh_sp/status/2074913240669429991
Closing the Verification Loop
https://thinkroom.kieranklaassen.com/d/njrS5TJhis
I’ll be at the poster today in Hall A, 5-7:45pm KST. Come say hi if you want to talk about synthetic data, self-play, or RL! Joint work w/ wonderful collaborators @zeynebnkaya @ShirleyYXWu @tengyuma @tatsu_hashimoto @lschmidt3 Paper:”
https://x.com/carolineschoi/status/2074726511396151355
NEW AI paper worth bookmarking. This is something I called early, and this paper confirms it: verification has emerged as a new important scaling axis. Here is the simple explainer and what this paper shows. We have seen lots of progress in scaling pre-training, post-training,”
https://x.com/omarsar0/status/2074556579580711050
SGLang now supports DSpark, enabling confidence-driven, variable-length verification for speculative decoding 🎉 DSpark addresses a key bottleneck under load: instead of verifying every draft token, it verifies only where the draft model is confident, so the gains hold even as”
https://x.com/lmsysorg/status/2074176669108367549
Understanding the Dynamics of the AI Ecosystem with Pace Layers
https://www.dbreunig.com/2026/07/03/ai-ecosytem-pace-layers.html
What is Neurosymbolic AI? Your Ultimate Guide”
https://x.com/TheTuringPost/status/2075043492263244118
Working with the @liquidai team on these engineering blogs is just so much fun! Here’s what we’ve been working on: Reasoning models can get stuck mid-thought. That’s what we call a “doom loop”. A doom loop happens when the model repeats the same token (“Wait”, “Let me”
https://x.com/helloiamleonie/status/2074498103982408044
Leave a Reply