Image created with OpenAI GPT-Image-1. Image prompt: rich crimson, bright ivory, deep navy Independence-Day palette, vibrant, celebratory, wholesome, authentic, photorealistic sunset harvest field with tractor flying a flag scene featuring an LED marquee scrolling “Cutting Edge Tech”; natural lighting, subtle film grain, high detail
if you block AI from accessing your content, there will be no one left to read it. to elaborate: the future of search is lightweight research agents. a lot of businesses are already seeing significant traffic come from ChatGPT. if you block AI scrapers, i will never read your content because o4-mini-high will send me to your competitor’s website”” / X https://x.com/vikhyatk/status/1940227029389255109
The race for LLM “cognitive core” – a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing.
Its features are slowly crystalizing:
– Natively multimodal text/vision/audio at both input and output.
– Matryoshka-style architecture allowing a dial of capability up and down at test time.
– Reasoning, also with a dial. (system 2)
– Aggressively tool-using.
– On-device finetuning LoRA slots for test-time training, personalization and customization.
– Delegates and double checks just the right parts with the oracles in the cloud if internet is available.
It doesn’t know that William the Conqueror’s reign ended in September 9 1087, but it vaguely recognizes the name and can look up the date. It can’t recite the SHA-256 of empty string as e3b0c442…, but it can calculate it quickly should you really want it.
What LLM personal computing lacks in broad world knowledge and top tier problem-solving capability it will make up in super low interaction latency (especially as multimodal matures), direct / private access to data and state, offline continuity, sovereignty (“not your weights not your brain”). i.e. many of the same reasons we like, use and buy personal computers instead of having thin clients access a cloud via remote desktop or so.https://x.com/karpathy/status/1938626382248149433
Amazon deploys over 1 million robots and launches new AI foundation model https://www.aboutamazon.com/news/operations/amazon-million-robots-ai-foundation-model
Many firms built around the limitations & cost assumptions of GPT-3.5 class models, and are now stuck with complex solutions that are more expensive & worse than a reasoner without any scaffolding You need to build solutions with an eye towards riding the cost/performance curve.”” / X https://x.com/emollick/status/1939494862438453717
Huawei announces open-sourcing of Pangu models to accelerate AI application, value creation https://www.ecns.cn/news/sci-tech/2025-07-01/detail-ihesxvny3991876.shtml
To supercharge its AI push without triggering a government review that would come with acquiring other companies, Meta bought a 49 percent non-voting stake in data-labeling firm Scale AI for $14.3 billion and hired its founder and CEO, Alexandr Wang, and key staff. Wang will”” / X https://x.com/DeepLearningAI/status/1940153434671362268
Tencent released Hunyuan-A13B, a new open-source hybrid reasoning model It nears or matches models like o1 and DeepSeek R1 on major benchmarks, while remaining efficient enough to run on a single GPU Also includes “”fast and slow”” modes to adjust efficiency levels https://x.com/rowancheung/status/1939601169271197973
The real star of the show is the (Baidu) 21B A3B, ~30% smaller than Qwen3 30B A3B and better on most benchmarks! 🔥 https://x.com/reach_vb/status/1939584854045466791
Earlier this month, we launched the Image Edit Arena. Today, the Image Edit Leaderboard 🏆 goes LIVE, powered by more models and all your community votes. 🏆 In 1st place: GPT-Image-1 by @OpenAI 💠 2nd-4th: Flux 1 Kontext Max, Pro & Dev by @bfl_ml 💠 5th: Gemini 2.0 Flash https://x.com/lmarena_ai/status/1940795298449924220
Practical Techniques for Context Engineering 💡 This is a fantastic blog post from @tuanacelik and @LoganMarkewich on a comprehensive breakdown of the types of context an LLM can interact with, and the core dimensions you have to consider: 1️⃣ Knowledge Base or tool selection – https://x.com/jerryjliu0/status/1940852245450608646
Code agents underperform because public SWE corpora are small, poorly verified, and cover few repositories. Skywork‑SWE builds a rigorously tested dataset of 10,169 Python bug‑fix tasks and fine‑tunes a 32B code agent. Performance climbs with data, reaching 38% pass\@1 and 47% https://x.com/rohanpaul_ai/status/1939616650002878712
RT @omarsar0: Evaluating LLM-based Agents This report has a comprehensive list of methods for evaluating AI Agents. Don’t ignore evals.…”” / X https://x.com/omarsar0/status/1940009835342246277
What happens if you put a full scientific paper into AI and ask it to find known errors in proofs, tables, etc? Every model before o3 fails completely, o3 gets 21% (its better at proofs, worse at tables & figures). Progress & perhaps a second opinion, not yet autonomous science. https://x.com/emollick/status/1940064995439779962
RT @Michael_D_Moor: 🚨Preprint alert!🚨 Did you know there is a new reasoning benchmark where leading models like o3 still fall flat? (i.e.…”” / X https://x.com/_akhaliq/status/1940066518307381616
🚀 Excited to share that MiniMax-M1 is now live on the Text Arena leaderboard at #12! And has climbed to #1 in math! 🎉 As the open-source hybrid MoE model with the world’s longest context window, we’re pushing boundaries in long-context reasoning and agentic tool use.💪”” / X https://x.com/MiniMax__AI/status/1940243199500677218
[2304.02868] Can Large Language Models Play Text Games Well? Current State-of-the-Art and Open Questions https://arxiv.org/abs/2304.02868
Context Engineering “”Context engineering”” is a key part of agent building. We review a few popular patterns for context engineering + explain how to use them w/ LangGraph. Blog: https://x.com/LangChainAI/status/1940440271126438118
Ever wondered how AI agents evolved from simple text models to full-fledged task performers? Here’s a breakdown of AI Agent’s journey from basic LLMs to memory-enabled, self-learning agents. Here’s how the evolution took place: 1. Basic LLM Agent Starts with a plain language https://x.com/goyalshaliniuk/status/1938122558043283720
Mem0 – The Memory Layer for your AI Apps https://mem0.ai/
RT @ChengleiSi: Are AI scientists already better than human researchers? We recruited 43 PhD students to spend 3 months executing research…”” / X https://x.com/tatsu_hashimoto/status/1939708064619475161
A foundation model to predict and capture human cognition | Nature https://www.nature.com/articles/s41586-025-09215-4
The transition from SLURM to K8s has been very painful already as k8s seems to ignore 50 years of Unix, with many of its great features stripped and now this surprise with B200 AWS nodes https://x.com/StasBekman/status/1940633288152174908
Interestingly Claude-4 Opus and o3 are on the same level on METR when selecting 80% (instead of 50%) probability of succeeding on a task https://x.com/scaling01/status/1940093773440008512
Claude 4 Opus and Claude 4 Sonnet fall behind o3 on METR”” / X https://x.com/scaling01/status/1940089136104579515
Digital Salon: An AI and Physics-Driven Tool for 3D Hair Grooming and Simulation https://digital-salon.github.io/
LLMs still miss subtle engineering tweaks humans exploit for major speed gains. 🚀 The Automated LLM Speedrunning Benchmark shows that today’s frontier models still miss many record‑setting NanoGPT tweaks, so automated reproducibility of AI research is far from solved. 🔍 https://x.com/rohanpaul_ai/status/1940288577700503584
New IIT-JEE rankings just dropped https://x.com/bilawalsidhu/status/1939901363162349629
Test-based certification is the only way forward in food, eager to see more over time. Food is not simple anymore – it is a complex, industrial product with global supply and processing chains. Contamination can be introduced in many stages along the way from farming to harvest,”” / X https://x.com/karpathy/status/1940181840201228384
Love this project: nanoGPT -> recursive self-improvement benchmark. Good old nanoGPT keeps on giving and surprising 🙂 – First I wrote it as a small little repo to teach people the basics of training GPTs. – Then it became a target and baseline for my port to direct C/CUDA”” / X https://x.com/karpathy/status/1939709449956126910
The study asks if models like GPT‑4o truly understand images. It finds they juggle many jobs yet still trail task‑focused vision tools. Past tests could not match chat models with pixel specialists fairly. The authors turn every benchmark into quick yes‑no image checks any API https://x.com/rohanpaul_ai/status/1941086082679951554
MAI-DxO in action, tackling one of those complex cases: https://x.com/mustafasuleyman/status/1939670348330619278
AllenAI just introduced SciArena in which o3 is crushing all other models”” / X https://x.com/scaling01/status/1940065085776666679
SciArena: A New Platform for Evaluating Foundation Models in Scientific Literature Tasks | Ai2 https://allenai.org/blog/sciarena
Many of the academic papers shared on X are benchmarking papers, which are made so that current AIs will fail often (or it isn’t a benchmark for future progress) You should pay attention to the realism of the benchmark, relative rankings, and the prompts & tools given to the AI.”” / X https://x.com/emollick/status/1940100137335902627
LLMs can synthesize many programs, but how should we search among them? New from @SakanaAILabs – AB-MCTS frames code generation as an adaptive tree search, guided by external feedback. Beats baselines on synthesis benchmarks including ARC-AGI.”” / X https://x.com/ndea/status/1940166177424384354
We’re excited to introduce AB-MCTS! Our new inference-time scaling algorithm enables collective intelligence for AI by allowing multiple frontier models (like Gemini 2.5 Pro, o4-mini, DeepSeek-R1-0528) to cooperate. Blog: https://x.com/SakanaAILabs/status/1939854145856708910
This paper enhances open-source LLMs by developing a data synthesis methodology focused on strategic planning and high-quality training data. The 7B model shows substantial performance improvements, while the 14B model achieves results comparable to or surpassing GPT-4o. https://x.com/rohanpaul_ai/status/1939211732200882534
M4THYOU/TokenDagger: High-Performance Implementation of OpenAI’s TikToken. https://github.com/M4THYOU/TokenDagger
METR results for DeepSeek V3 and R1 kinda sucks, huh? https://x.com/scaling01/status/1939770925781487779
🔑 Gradient flow can turn ordinary neural weights into exact symbolic rules when the model respects geometric symmetry. The paper proves that training then lands on a low entropy ring of weight distributions that behave like logic circuits. Main takeaway: geometric constraints https://x.com/rohanpaul_ai/status/1941070731535585404
3rd point: decision transformer doesn’t work; online RL is needed for true self-improvements and AI co-scientists”” / X https://x.com/shaneguML/status/1939767338553004518
All the technical language around AI obscures the fact that there are two paths to being good with AI: 1) Deeply understanding LLMs 2) Deeply understanding how you give people instructions & information they can act on. LLMs aren’t people but they operate enough like it to work”” / X https://x.com/emollick/status/1939045794054541574
ARIG: Autoregressive Interactive Head Generation for Real-time Conversations
https://jinyugy21.github.io/ARIG/
cool new paper/idea but a huge missed opportunity to name the model 5TPG https://x.com/jxmnop/status/1940772450696155528
DiLoCoX proposes a low-communication framework for decentralized cluster training. This framework significantly improves pre-training speed and expands parameter scale. It enables pre-training a 107 billion parameter model over a 1 Gigabit per second network, achieving a 357 https://x.com/rohanpaul_ai/status/1939150579596701826
Fine tuning often stops a model from saying “I don’t know” when it should. SEAT trains only a few weights and adds a small regularizer so the model can learn fresh facts while keeping that honest silence. People like short chats with honest models. SEAT keeps that honesty https://x.com/rohanpaul_ai/status/1938960155749818668
i just need uv to become a part of standard python i never want to have to give pip install instructions again just use uv”” / X https://x.com/qtnx_/status/1940025303289495898
I’m excited about model diffing as an agenda, it seems like we should be so much easier to look for alignment relevant properties if we can ignore everything also in the base model. It’s great to see signs of life!”” / X https://x.com/NeelNanda5/status/1939798682234495429
Inference-Time Scaling and Collective Intelligence for Frontier AI https://x.com/hardmaru/status/1939866376988143687
Inference-Time Scaling and Collective Intelligence for Frontier AI https://sakana.ai/ab-mcts/
LLM reasoning with reinforcement learning focuses on limited domains, hindering general applicability. This paper develops GURU, a 92,000-example multi-domain dataset, to enable broader reinforcement learning-based reasoning. Methods 🔧: – GURU includes Math, Code, Science, https://x.com/rohanpaul_ai/status/1939196632643551542
LLM swarms trade speed for richer reasoning in group decision making. Using Boids and ant colony examples, the paper asks if swapping tiny rule sets for large language model prompts still counts as swarm intelligence, and finds that LLM agents reproduce flocking and path finding https://x.com/rohanpaul_ai/status/1940239756295569763
LLMs sway people in debate yet fail to grasp what they say, the paper shows . Adding formal debate rules boosts persuasiveness but still leaves comprehension shallow. Key finding: persuasion does not need true understanding. LLMs are rapidly used as judges and helpers, so the https://x.com/rohanpaul_ai/status/1941040029104853015
Ship your research. https://x.com/LaudeInstitute/status/1937681620028600529
Swap “”prompt engineering”” with “”context engineering”” and “”hallucination”” with “”confabulation”” and we’ll have eliminated like a quarter of the cringe debt from the early LLM era.”” / X https://x.com/jd_pressman/status/1939725776481656886
That’s what every technical report should look like https://x.com/scaling01/status/1939715730217308420
The State of AI in 2025 https://www.iconiqcapital.com/growth/reports/2025-state-of-ai
This paper proposes Enigmata, a comprehensive suite providing synthetic, verifiable puzzles and optimized Reinforcement Learning with Verifiable Rewards training, to enhance LLMs’ logical reasoning. Methods 🔧: – Enigmata-Data offers 36 puzzle tasks across seven categories, https://x.com/rohanpaul_ai/status/1939165930438992142
What is context Engineering? “Context Engineering is the discipline of designing and building dynamic systems that provides the right information and tools, in the right format, at the right time, to give a LLM everything it needs to accomplish a task.” Read it: https://x.com/_philschmid/status/1940692654284505391
X-UniMotion https://byteaigc.github.io/X-Unimotion/
You need credentials and a PhD is a standard one. But those days a single impactful project on GitHub can indeed go a long way.”” / X https://x.com/francoisfleuret/status/1939181398163898458
maciej-trebacz/tower-of-time-game: Vibe coded Tower Defense type of game made for a game jam https://github.com/maciej-trebacz/tower-of-time-game
SynMotion https://lucaria-academy.github.io/SynMotion/
A challenge with AI adoption is that organizations are not built to a Grand Plan where AI can just be slotted in, but rather socially constructed, random & in flux Here’s an anecdote from a paper on how a process re-engineering effort led to revelations that drove people insane. https://x.com/emollick/status/1940170621230752031
Introducing AlphaGenome: an AI model to help scientists better understand our DNA – the instruction manual for life 🧬 Researchers can now quickly predict what impact genetic changes could have – helping to generate new hypotheses and drive biological discoveries. ↓ https://x.com/GoogleDeepMind/status/1937873589170237738
🚀 Sparse Neural Retrievers in Sentence Transformers v5 @huggingface just released version 5 of Sentence Transformers, with full support for training and fine-tuning sparse neural retrievers, so you can bring your hybrid search to the next level. This release gives you https://x.com/qdrant_engine/status/1940052377039413474
pip is to uv what Edge is to Chrome”” / X https://x.com/hkproj/status/1940026008591106479
