OpenAI's updated Advanced Voice is great for translating conversations between people speaking different languages.

French company Mistral AI announced Magistral, their first reasoning model designed to excel in domain-specific, transparent, and multilingual reasoning.

Read more: https://mistral.ai/news/magistral
Try it here: https://chat.mistral.ai/chat

An important milestone towards bringing humanoid robots to the home: 1X AI has trained Redwood, a VLA capable of end-to-end mobile manipulation tasks like retrieving objects for users, opening doors, and navigating around the home.

Read more: https://www.1x.tech/discover/redwood-ai

A fully open-source system for humanoid whole-body teleoperation: https://zzk273.github.io/R2S2/

Code + models + hardware specs: https://github.com/GalaxyGeneralRobotics/OpenWBT

New blog post from Sam Altman.

'We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence'

https://blog.samaltman.com/the-gentle-singularity

Algorithmic progress keeps hitting hard. Models are not just getting smarter but a lot cheaper as well.

Although people claim that AI is overhyped, in my experience, more people are actively misinformed by anti-AI influencers.

For example, I regularly encounter people who still believe self-driving cars have failed, despite the fact that there are now hundreds of thousands of paid driverless trips per week in the United States and China.

Most people have never heard of AI systems like AlphaChip, which Google used to design superhuman chip layouts, or AlphaEvolve, which discovered novel algorithms and helped Google save money by improving its global data center capacity.

Even someone like Sam Altman predicts that "2027 may see the arrival of robots that can do tasks in the real world," yet Figure AI already employs humanoid robots for real-world logistics tasks, as well as at BMW car production plants.

Even when you show people all of this, they often point out trivial shortcomings, such as the fact that robots are still slower than human workers. No one makes the slightest effort to extrapolate from this. Rather than considering where progress will be two more papers down the line, people imagine that the world will remain frozen at its current level of capabilities, despite the rapid progress of the last ten years.

The general attitude toward AI is: Don't look up! What you can't see can't hurt you.

Links for 2025-06-11

AI

1. META is forming a dedicated Superintelligence lab headed by Alexandr Wang. https://www.cnbc.com/2025/06/10/meta-scale-ai-alex-wang.html

2. Meta took a stake in a major artificial intelligence data firm, making one of its biggest ever deals as it fights to catch up with rivals in the AI race. https://www.washingtonpost.com/technology/2025/06/10/ai-meta-scale-google-openai/ [no paywall: https://archive.is/OBVAM]

3. V-JEPA 2: A new world model with state-of-the-art performance in visual understanding and prediction. https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/

4. RewardAnything: Generalizable Principle-Following Reward Models https://arxiv.org/abs/2506.03637

5. Reinforcement Pre-Training https://arxiv.org/abs/2506.08007

6. Human intention inference with a large language model can enhance brain-computer interface control: A proof-of-concept study https://www.biorxiv.org/content/10.1101/2025.06.01.657282v2

7. StochasTok: Improving Fine-Grained Subword Understanding in LLMs https://arxiv.org/abs/2506.01687

8. AI-enabled control system helps autonomous drones stay on target in uncertain environments https://news.mit.edu/2025/ai-enabled-control-system-helps-autonomous-drones-uncertain-environments-0609

9. OpenAI taps Google in unprecedented cloud deal despite AI rivalry https://www.reuters.com/business/retail-consumer/openai-taps-google-unprecedented-cloud-deal-despite-ai-rivalry-sources-say-2025-06-10/

10. OpenAI hits $10 billion in annual recurring revenue https://www.cnbc.com/2025/06/09/openai-hits-10-billion-in-annualized-revenue-fueled-by-chatgpt-growth.html

11. F.D.A. to Use A.I. in Drug Approvals to ‘Radically Increase Efficiency’ https://www.nytimes.com/2025/06/10/health/fda-drug-approvals-artificial-intelligence.html [no paywall: https://archive.is/W36rN]

12. Ilya Sutskever, U of T honorary degree recipient, June 6, 2025 https://www.youtube.com/watch?v=zuZ2zaotrJs

13. UK government harnesses Gemini to support faster planning decisions https://blog.google/around-the-globe/google-europe/united-kingdom/uk-government-harnesses-gemini-to-support-faster-planning-decisions/

14. Why are neural networks and cryptographic ciphers so similar? https://reiner.org/neural-net-ciphers

According to Anthropic co-founder Ben Mann, we'll know AI is transformative when it passes the "Economic Turing Test."

Assign an AI agent a task to complete over the course of a month. Then, let the hiring manager choose between a human and a machine.

If they pick the machine more often than not, we've crossed the threshold.

Original source: https://youtu.be/aStf54Vxy24?si=0Ek00ZsGMPKQGPUC&t=895

A mathematical paper has been autoformalized for the first time: https://www.morph.so/blog/trinity

Trinity: An autoformalization system that represents a critical step toward verified superintelligence.

Autoformalization will create a nearly infinite supply of verified training environments. By converting mathematical knowledge into formal proofs, Trinity will generate an endless curriculum of verified challenges—enabling true self-supervised reinforcement learning in mathematics.

Trinity systematically processes entire papers, intelligently corrects its own formalization errors by analyzing failed attempts, and automatically refactors lengthy proofs to extract useful lemmas and abstractions. This results in independently verifiable mathematical knowledge that requires no trust in the AI system itself.

Preliminary information suggests that Israel has achieved total air superiority over Iran, with Israeli surveillance drones flying freely over Iranian cities.

If this is true, the Mullah regime is essentially finished.

And what's Iran's ally Russia doing? They are too scared of Israel and don't have the strength to intervene anyway.

What if an LLM could update its own weights?

Meet SEAL🦭: a framework where LLMs generate their own training data (self-edits) to update their weights in response to new inputs.

Self-editing is learned via RL, using the updated model’s downstream performance as reward.

Read more: https://jyopari.github.io/posts/seal

Comment on The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity https://arxiv.org/abs/2506.09250

Anthropic introduces Claude’s “Research” mode: a LeadResearcher Claude Opus 4 agent plans a strategy, launches task-specific Claude Sonnet 4 sub-agents to run parallel searches across the web, Google Workspace and other integrated tools, then reunites their findings with automatic source attribution. This orchestrator–worker pattern lets Claude tackle breadth-first, context-spanning questions that stump single agents. In internal BrowseComp-style tests, it answered research tasks correctly 90.2 % more often than a lone Claude Opus 4 agent, at about 15 × the token cost.

Read more: https://www.anthropic.com/engineering/built-multi-agent-research-system

Scaling Laws in Autonomous Driving

The core AI tasks behind “Waymo Driver”―motion forecasting and planning―obey clear scaling laws similar to those found in large-language models.

Image:

Model performance predictably improves as a function of the training compute budget. This predictable improvement not only applies to the objective the model is trained with (Left), but also to popular motion forecasting open-loop metrics (Middle), and most importantly, to planning performance in closed-loop simulation (Right).

Read more: https://waymo.com/blog/2025/06/scaling-laws-in-autonomous-driving

Links for 2025-06-16

AI

1. A Statistical Physics of Language Model Reasoning https://arxiv.org/abs/2506.04374

2. Reinforced Meta-thinking Agents (ReMA) combines meta-learning and reinforcement learning (RL) to increase effectiveness of LLMs https://www.turingpost.com/p/metalearning

3. Unsupervised pre-training for RL: learn a flow-based future prediction model for each "intention" in the dataset. https://chongyi-zheng.github.io/infom/

4. Few-step generation in discrete diffusion language models via the underlying Gaussian diffusion. https://s-sahoo.com/duo/

5. e3: Learning to Explore Enables Extrapolation of Test-Time Compute for LLMs https://matthewyryang.com/e3/

6. LLMs beating trained humans in legal reasoning. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5283722

7. AI systems can find real vulnerabilities in widely-used software https://arxiv.org/abs/2506.02548

8. Leading LLM agents achieve modest overall success rates on CRMArena-Pro. Reasoning models exhibit markedly superior performance relative to non-reasoning ones. https://arxiv.org/abs/2505.18878

9. How Do Olympiad Medalists Judge LLMs in Competitive Programming? “For game theory, greedy, ad-hoc and constructive problems, which usually require significant amounts of observations and LLMs often struggle with… even reasoning brings minimal improvement.” https://arxiv.org/abs/2506.11928

10. “We break the grid-world barrier with PoE-World, a program synthesis world modeling method which represents a world model as an exponentially-weighted product of programmatic experts synthesized by LLMs.” https://topwasu.github.io/poe-world

11. Reviving DSP for Advanced Theorem Proving in the Era of Reasoning Models https://arxiv.org/abs/2506.11487

12. Improving LLM Agent Planning with In-Context Learning via Atomic Fact Augmentation and Lookahead Search https://arxiv.org/abs/2506.09171

13. GUI-Reflection: Empowering Multimodal GUI Models with Self-Reflection Behavior https://penghao-wu.github.io/GUI_Reflection/

14. Global Convergence of Gradient EM for Over-Parameterized Gaussian Mixtures https://arxiv.org/abs/2506.06584

15. If China wins the AI race https://www.foreignaffairs.com/united-states/what-if-china-wins-ai-race [no paywall: https://archive.is/mKKEt]

16. Countering AI Chip Smuggling Has Become a National Security Priority https://www.cnas.org/publications/reports/countering-ai-chip-smuggling-has-become-a-national-security-priority

17. BT chief eyes deeper job cuts as AI becomes more powerful https://www.ft.com/content/c8d41424-f2d9-418e-8da2-55823f1c22ca [no paywall: https://archive.is/y7PHu]

18. Mattel and OpenAI are teaming up on AI powered toys https://openai.com/index/mattels-iconic-brands/

19. How LLM Beliefs Change During Chain-of-Thought Reasoning https://www.lesswrong.com/posts/GwvWtAwnKBKjmknag/how-llm-beliefs-change-during-chain-of-thought-reasoning-2

20. The Claude Bliss Attractor https://www.astralcodexten.com/p/the-claude-bliss-attractor

21. You need to give o3 Pro a ton of context to see its intelligence, but when you do it creates incredibly specific, actionable output. https://www.latent.space/p/o3-pro

Miscellaneous

1. Endometriosis is an incredibly interesting disease https://www.lesswrong.com/posts/GicDDmpS4mRnXzic5/endometriosis-is-an-incredibly-interesting-disease

2. "On average, [newborn] males had significantly larger intracranial and total brain volumes [than newborn girls], even after controlling for birth weight... sex differences in brain structure are already present at birth." https://link.springer.com/article/10.1186/s13293-024-00657-5

3. Ultra-wide-field, deep, adaptive two-photon microscopy https://www.biorxiv.org/content/10.1101/2025.06.07.658419v1

4. A complementary two-dimensional material-based one instruction set computer https://www.nature.com/articles/s41586-025-08963-7

5. Balloons carrying radio antennas over Antarctica have detected mysterious radio pulses coming up from the ice https://www.psu.edu/news/research/story/strange-radio-pulses-detected-coming-ice-antarctica

1250 correlations between family members in traits, representing about 170,000,000 pairs of individuals:

https://www.sebjenseb.net/p/meta-analysis-of-1250-correlations

I've seen this movie before.

https://www.cnbc.com/2025/06/16/openai-wins-200-million-us-defense-contract.html

NATO: +2 members
+ Finland
+ Sweden
Casualties: 0

Russia: -2 allies
- Syria
- Iran
Casualties: >1 Million

Intelligence too cheap to meter:

Gemini 2.5 Flash-Lite is so fast, it codes *each screen* on the fly (Neural OS concept).

Read more: https://blog.google/products/gemini/gemini-2-5-model-family-expands/