R[AI]DAR

Category intelligence

Social Media Briefing

All 150 current items, analyzed and ranked.

Executive synthesis

Social Media Summary

Executive Signal

  • Open-source AI influence is tilting decisively toward Chinese ecosystems while security, capability, and embodiment debates expose real gaps between hype and deployable systems.

Priority Developments

  • Citation dominance shifting east: Qwen-family models now appear in ~40% of post-ChatGPT arXiv AI/ML papers versus 25-30% for US open models, signaling a structural shift in research gravity.
  • Pretrained-friendly architectural gains: DeepMind/UT Austin's recirculation modification delivers 23% perplexity improvements without weight retraining, compressing the cost curve for capability uplift.
  • Agent readiness overstated in regulated domains: MIT's FinanceGym shows leading models (Kimi K3, Qwen 3.8 Max) fail private-data finance benchmarks, exposing enterprise deployment gaps.
  • Security posture remains unresolved: Pieter Levels' prompt-injection walkthrough demonstrates that autonomous agent-to-tool connections remain trivially exploitable.
  • Geographic and talent consolidation: Mistral's HUMAIN partnership and Perplexity's hire of Andrew Wilson (continual learning, synthetic data, RL) accelerate non-US AI capacity buildouts.

Leadership Implications

Key Themes

US-China Open Model Competition · 1AI industry talent movement and research hiring · 2AI Research & Benchmarks · 3Prompt Injection and Security · 1Research Papers and Architectural Innovations · 4LLM education and transformer fundamentals · 1AI Infrastructure & Inference · 2LLM Limitations and Robotics · 1AI Agents and Frameworks · 4Sovereign AI and international partnerships · 1

Primary evidence

All Ranked Signals

Showing 10 of 150 signals 2 remaining
Social Twitter Aug 24

Over the weekend I had Codex parse 500K arXiv AI/ML papers since ChatGPT to understand which open mo...

By @natolambert

88 score
AI Analysis

AI researcher @natolambert shares results from a Codex-assisted analysis of 500K arXiv AI/ML papers tracking which open models are cited. Reports that Chinese open models (led by Qwen) now appear in ~40% of papers vs ~25-30% for American open models, and Qwen is cited in ~1/3 of LLM-mentioning papers. Llama peaked in April 2025 and has declined since Llama 4.

Source details
Over the weekend I had Codex parse 500K arXiv AI/ML papers since ChatGPT to understand which open models are used for research. In 2024, ~30% of papers mentioned an American open model and only 10% a Chinese model. Today, ~40% of papers mention a Chinese (open) LLM, and only 25-30% an American one. Chinese models are the default for research. Chinese mentions are still growing while American open models are stagnating. When looking at this data it's important to remember that papers substant
open modelsresearch trendsUS-China AI competitionQwenLlamabenchmarking
Social Twitter Aug 24

Tackling a 60-year-old challenge in quantum chemistry: making density functional theory scale nearly...

By @AnimaAnandkumar

88 score
AI Analysis

Anima Anandkumar presents research on a unified AI model that scales density functional theory quasi-linearly using a novel Fourier neural operator variant, enabling quantum-mechanical simulations of molecules and materials. The model uses intermediate reasoning-like steps to improve extrapolation.

Source details
Tackling a 60-year-old challenge in quantum chemistry: making density functional theory scale nearly linearly with system size. This has huge implications for opening the door to realistic systems that have traditionally been too expensive to simulate. AI has attempted to accelerate these calculations, but models generally struggle to extrapolate, particularly to systems larger than those seen during training. Unlike text or images, quantum-mechanical training data is extremely expensive to ge
scientific-AIquantum-chemistryneural-operatorsAI4Scienceresearch
Social Twitter Aug 24

Transformers process a prompt mostly in parallel, which is efficient but makes it hard for them to m...

By @burkov

82 score
AI Analysis

Technical deep dive on a Google DeepMind / UT Austin paper introducing 'recirculation' — a lightweight architectural modification that routes deeper-layer representations back into shallower layers at the next step, yielding 23% perplexity reductions on Gemma and improving instruction following and math reasoning without changing pretrained weights

Source details
Transformers process a prompt mostly in parallel, which is efficient but makes it hard for them to maintain an evolving internal state—roughly, a running representation of what the text currently implies—as new information arrives; this can lead to failures where a model forms the right interpretation in deeper layers but cannot make that interpretation available early enough for later processing. In this paper from @GoogleDeepMind and @UTAustin, the authors explore a small architectural chang
transformer architectureGoogle DeepMindGemmarecurrent mechanismslanguage modelingresearch papers
Social Twitter Aug 24

If you're 17 (or any age) and you want to learn to build LLMs from scratch, read chapters 15-16 of D...

By @fchollet

78 score
AI Analysis

François Chollet recommends chapters 15-16 of his Deep Learning with Python book for learners wanting to build LLMs from scratch, highlighting chapter 15's explanation of why dot-product attention works.

Source details
If you're 17 (or any age) and you want to learn to build LLMs from scratch, read chapters 15-16 of Deep Learning with Python, available online here: t.co/Nisfkzf9sC In particular, chapter 15 has one of the best explanations of WHY dot-product attention works that you'll find anywhere.
educationtransformer architectureattention mechanismlearning resources
Social Twitter Aug 24

I haven't directly connected my bug board https://t.co/UzcZKgDrPf to my AI because I am very aware o...

By @levelsio

78 score
AI Analysis

Pieter Levels explains why he manually reviews bug reports rather than connecting his AI directly to his bug board, walking through a concrete prompt-injection chain that hides a backdoor in an innocuous feature request.

Source details
I haven't directly connected my bug board t.co/UzcZKgDrPf to my AI because I am very aware of prompt injection risks One safe way people mention would be to only give the AI access to collect user bug reports and feature requests, then do pull requests on GitHub that I then review myself before I approve or reject them But you can imagine a benign attacker can prompt a feature request with an elaborate prompt that tells it to add a backdoor to my sites BUT then in that prompt write al
prompt-injectionsecurityai-agents
Social Twitter Aug 24

Excited to welcome Andrew Gordon Wilson to our research team. He will be reporting to Denis and lead...

By @AravSrinivas

75 score
AI Analysis

Continuing our coverage from yesterday, Aravind Srinivas announces Andrew Gordon Wilson's hire at Perplexity, listing research areas: continual learning, synthetic data, long-horizon RL environments, and architectures.

Source details
Excited to welcome Andrew Gordon Wilson to our research team. He will be reporting to Denis and lead new research efforts on continual learning, synthetic data, long horizon RL environments and architectures. We’re hiring! Please reach out to Andrew!
Perplexityresearch leadershipcontinual learningsynthetic datareinforcement learningindustry hiring
Social Twitter Aug 24

I am excited to announce that I am joining @perplexity_ai as research lead! We will be doing ambitio...

By @andrewgwils

73 score
AI Analysis

Continuing our coverage from yesterday, Andrew Gordon Wilson announces he is joining Perplexity AI as research lead to work on continual learning, agent collaboration, and paradigm-shifting research.

Source details
I am excited to announce that I am joining @perplexity_ai as research lead! We will be doing ambitious paradigm shifting work, advancing the frontiers in the open. If you want to join us in re-imagining continual learning, agent collaboration, and beyond, please reach out!
talent movementPerplexitycontinual learningagent researchindustry hiring
Social Twitter Aug 24

Training finance agents is hard: Private data is locked inside large corporations. MIT CSAIL resear...

By @MIT_CSAIL

72 score
AI Analysis

MIT CSAIL announces FinanceGym, described as the largest finance-agent dataset to date with 59K+ samples across 37 roles and 40 apps/sites. Notes that even top models like Kimi K3 and Qwen 3.8 Max failed the test set.

Source details
Training finance agents is hard: Private data is locked inside large corporations. MIT CSAIL researchers release "FinanceGym," the largest finance-agent dataset yet, w/59K+ samples across 37 roles and 40 apps/sites. Even Kimi K3 & Qwen 3.8 Max failed its test set. t.co/lCt5XpHZtN
benchmarksagentsfinance-AIdataset-releasemodel-evaluation
Social Twitter Aug 24

@danteblank_ia @paulg You can hook up a humanoid robot to an LLM running in a data center, and you s...

By @ylecun

72 score
AI Analysis

Yann LeCun replies that hooking a humanoid robot to a data-center LLM still would not yield a robot capable of cleaning a room.

Source details
@danteblank_ia @paulg You can hook up a humanoid robot to an LLM running in a data center, and you still wouldn't have a robot that can clean your room.
roboticsllm-limitationsworld-modelsjepa
Social Twitter Aug 24

Today, we’re announcing a strategic collaboration with @HUMAIN spanning AI infrastructure, advanced...

By @MistralAI

70 score
AI Analysis

Mistral AI announces a strategic collaboration with HUMAIN covering AI infrastructure, advanced model development, and deployment in Saudi Arabia, with initial focus on cybersecurity, voice, and Arabic-language models.

Source details
Today, we’re announcing a strategic collaboration with @HUMAIN spanning AI infrastructure, advanced model development, and the deployment of AI solutions in Saudi Arabia and across the region. Together, we will work on localized frontier AI models, with initial areas of focus including cybersecurity, voice, and models that perform strongly in Arabic languages. t.co/OTChHHutnJ
MistralHUMAINSaudi Arabiainternational expansionArabic language modelssovereign AIpartnerships