Kevin G. Yager | Academic Summary

AI News 2025-05-15

Posted on 2025-05-15 by KevinYager

General

The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis. ChatGPT, when used appropriately, can improve learning outcomes.
Analysis of AI usage: Large Language Models, Small Labor Market Effects. They find evidence of significant uptake/usage, but relatively minor effects on economic indicators.

Research Insights

Reinforcement Learning for Reasoning in Large Language Models with One Training Example. Training over and over on a single example can yield gains.
- Interesting to compare/contrast to RL on unlimited synthetic data: Absolute Zero: Reinforced Self-play Reasoning with Zero Data.
Sakana AI introduces Continuous Thought Machines (interactive report, preprint, code), a new neural approach where the neurons are synchronized and allowed to move their attention around over time. This allows a crude approximation of temporal “thinking” as the neurons modify their focus and state.

LLM

OpenAI add o4-mini to their reinforcement fine-tuning API.
ByteDance releases SeedCoder 8B.
OpenAI adds GPT-4.1 to the ChatGPT web product.
OpenAI release HealthBench. In addition to providing a useful way to track progress on LLMs for healthcare applications, the current results demonstrate just how effective existing LLMs can be in this application space.

Agents

Google describes: AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms (paper). An evolutionary approach is used to optimize algorithms for tasks including math, chip design, datacenter scheduling, and fine-tuning LLMs.
- Among many other things, it was able to improve on various optimal packing math problems.

Safety

OpenAI launch safety evaluations hub.

Audio

Stability AI and Arm release Stable Audio Open Small (weights).

Video

Google announces that the latest improvements to Gemini 2.5 (Pro and Flash) greatly improve video understanding.
Tencent HunyuanCustom is an open-source video model (preprint, code).
New lipsync model: KeySync: A Robust Approach for Leakage-free Lip Synchronization in High Resolution (preprint, code, demo).
LTXV 13B Distilled claims rendering 5× faster than base model.

World Synthesis

Enigma Labs claims they have made the first multiplayer AI-generative video game (a multiplayer car racing game). They say they will open-source the work eventually. Although the gameplay video shows crude graphics, it is further evidence that generative environments are a key part of future entertainment.

Science

AI-Engineered DNA Turns Genes On and Off in Blood Cells.
- Publication in Cell: Design principles of cell-state-specific enhancers in hematopoiesis.
Inference-time compute methods being applied to AI for bioscience: Test-Time Scaling Unlocks a Leap Forward in De Novo Antibody Design.
- Preprint: De novo design of hundreds of functional GPCR-targeting antibodies enabled by scaling test-time compute.
Behind the Noise (github). Self-supervised denoising can learn meaningful representations for science images.
- Preprint: Behind the Noise: Conformal Quantile Regression Reveals Emergent Representations.
Meta report: Sharing new breakthroughs and artifacts supporting molecular property prediction, language processing, and neuroscience.
- Open Molecules 2025 (OMol25) and Meta’s Universal Model for Atoms (UMA): Revolutionizing design at the atomic scale
- Adjoint Sampling: A breakthrough in highly scalable, reward-driven generative modeling
- Unlocking how the human brain develops language

Hardware

Google announces: Gemini smarts are coming to more Android devices. E.g. Gemini will be available from one’s smart watch. A Samsung headset with Gemini integration will supposedly launch later this year.

Robots

Tesla shows a video of Optimus robot dancing. Fluid motion like this tests the limit of hardware and software (latency, real-time compensation, etc.).

Posted in AI, News | Tagged hardware, LLM, research, robots, safety, Science, video, world synthesis | Leave a comment

AI News 2025-05-08

Posted on 2025-05-08 by KevinYager

General

OpenAI is purchasing Windsurf (who have produced an AI-coding editor) for $3B.
AI usage: ChatGPT Is Still Leading the AI Wars but Google Gemini Is Gaining Ground.

Research Insights

Layers at Similar Depths Generate Similar Activations Across LLM Architectures. Fascinating to see such strong similarities between LLMs trained by different groups on different datasets.
Absolute Zero: Reinforced Self-play Reasoning with Zero Data (preprint, code, models). Models learn to propose tasks that improve learning, and then solve those tasks. This approach thus claims the models can iteratively improve through self-play, providing hope that RL-like methods can be applied even to domains without verification.

LLM

Anthropic announces Integrations for Claude, allowing connection to tools like Zapier, Stripe, etc. This includes the ability to build connections to custom tools.
Mistral announce Mistral Medium 3.
A quick-and-dirty benchmark obtained by averaging 28 popular benchmarks (updated version). The leaders are Gemini 2.5 Pro, o3, and Claude 3.7 Thinking.

Audio

xAI adds voice mode to Grok.
Nvidia open-sources Parakeet TDT 0.6B V2, a fast audio transcription model.

Video

LTX Studio announce open-source video model: LTX-Video 13B (code, examples, more examples).
HeyGen announces Avatar IV (example, another example). Simpler usage and more expressive output.

Brain

Google announce: A new light on neural connections. They use light microscope to map neuronal connections, bypassing the need for more expensive/complicated imaging (e.g. electron microscopy). Their method improves upon existing expansion microscopy (ExM), wherein tissue is infused and stretched to expand the sizescale of small structures, using chemical labeling to highlight relevant proteins and deep learning to do image processing. This should increase the rate at which we can image connectomics in neuronal tissue.
- Nature publication: Light-microscopy-based connectomic reconstruction of mammalian brain tissue.

Robots

π0.5: a Vision-Language-Action Model with Open-World Generalization. The claim is that this model enables a robot to handle never-seen real-world environments.
VideoMimic: Visual imitation enables contextual humanoid control (preprint). Shows good progress in having robots learn motion policies from monocular video data.

Posted in AI, News | Tagged audio, brain, LLM, research, robots, video | Leave a comment

AI News 2025-05-01

Posted on 2025-05-01 by KevinYager

General

Luke Drago and Rudolf Laine provide an update to their predictions/proposals for transformative AI: The Intelligence Curse (PDF version). They describe both the danger in making the contributions of people economically worthless, as well as possible remedies by centering humans.
Dario Amodei publishes an opinion: The Urgency of Interpretability. He argues that mechanistic interpretability of LLMs is making significant progress, and that this is the key to future AI capabilities and safety.
A cheat sheet for why using ChatGPT is not bad for the environment.
Introducing the Anthropic Economic Advisory Council.

Research Insights

Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory.
The Leaderboard Illusion. They analyze the chatbot arena evaluation method, pointing out shortcomings and suggesting improvements.
Towards Understanding the Nature of Attention with Low-Rank Sparse Decomposition (code). The results suggest that attention heads also exhibit superposition.

LLM

Alibaba announce Qwen3 0.6B to 235B (code, weights, modelScope). Combines normal and thinking/reasoning into a single model.
LlamaCon 2025 announcements include: a new Meta AI app (with social features), safeguarding tools (Llama Guard 4 12B, Llama Firewall, Prompt Guard), collaboration with Groq and Cerebras for inference compute.
Phi-4 14B Reasoning (technical report).
Meta ReasonIR 8B is a model optimized for retrieval.

Safety

Contemplative Wisdom for Superalignment. They describe an alignment philosophy that is a s sort of “benevolent jailbreak” and leverages contemplative consideration.
Researchers Secretly Ran a Massive, Unauthorized AI Persuasion Experiment on Reddit Users. Here is an extended abstract for the study: Can AI Change Your View? Evidence from a Large-Scale Online Field Experiment.
Scaling Laws for Scalable Oversight (preprint, code).

Audio

Google DeepMind updates: Music AI Sandbox, now with new features and broader access.
Tavus announces Hummingbird-0 lipsync model (try on FAL).
Google NotebookLM adds “select output language” to their AI audio podcasts feature.
Suno releases v4.5 improved music generation.

Image Synthesis

Freepik and Fal announce F-Lite (tech report), an open-source image model (10B, trained on 80M images).
Midjourney has pushed updated to the v7 model (improving quality and coherence), adds an experimental aesthetic intensity parameter, and launches a new omni-reference feature (example outputs).

Video

Runway roll out their references feature to all paying users, which allows one to include specific characters/environments/elements in generations.

Science

FutureHouse Platform launches, providing access to their AI Scientist agent.

Robots

Deep Robotics shows a video of Lynx M20 wheeled-quadruped.
New video of LimX Dynamics CL-3.

Posted in AI, News | Tagged audio, image synthesis, LLM, research, robots, safety, video | Leave a comment

AI News 2025-04-24

Posted on 2025-04-24 by KevinYager

General

Richard Sutton and David Silver publish an essay: Welcome to the Era of Experience. They argue that scaling AI on human data (pre-training) is hitting diminishing returns. The resurgence of reinforcement learning presages a shift towards experiential learning, where agents will need to act in the world to self-train.
Anthropic publishes: Values in the wild: Discovering and analyzing values in real-world language model interactions (paper). In real conversations with users, Claude broadly adheres to the values intended by Anthropic. This approach can be used to help confirm alignment after deployment.
About a month ago, METR released a paper analyzing scaling of AI capabilities: Measuring AI Ability to Complete Long Tasks. Several follow-up analyses agree that the extrapolated trend suggests rapid improvements (though there are large uncertainties).
- AI Digest: A new Moore’s Law for AI agents.
- Forecaster reacts: METR’s bombshell paper about AI acceleration. New data supports an exponential AI curve, but lots of uncertainty remains.
Dwarkesh Patel: Questions about the Future of AI.
Epoch AI releases a report on: Trends in AI Supercomputers.

Research Insights

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? Results suggest that the base model already has reasoning capabilities. RL allows one to elicit these, but does not generate this capability per se. This is intuitive in the sense that base models can be manually used to do chain-of-thought through prompt engineering or re-prompting. RL thus allows a more efficient version of this behavior to be elicited.
Sleep-time Compute: Beyond Inference Scaling at Test-time. One can pre-compute a set of plausible questions/answers offline. By pre-computing useful components of user queries, inference-time compute can be saved.
Artificial intelligence and dichotomania. LLMs exhibit humanlike biases regarding interpreting statistical significance. This acts as further confirmation that pretraining is building models that implicitly simulate human behavior.
NEMOTRON-CROSSTHINK: Scaling Self-Learning beyond Math Reasoning. They improve generalized reasoning by incorporating multi-domain, multi-format RL.
Think Deep, Think Fast: Investigating Efficiency of Verifier-free Inference-time-scaling Methods. They find that reasoning models are superior to inference-time scaling of non-reasoning models.

LLM

Pleias-RAG are small models (350M, 1.2B) optimized for RAG (paper).

AI Agents

Anthropic posts: Claude Code: Best practices for agentic coding.

Audio

Nari Labs Dia is a text-to-speech (TTS) model that can generate remarkably realistic and emotional output (example).

Video

Skyworks releases SkyReels V2, an open-source video generator that can handle unlimited duration (examples).
Sand AI introduces Magi-1, an open-source autoregressive diffusion video model, that also handles infinite extension (examples).

Hardware

Google demos next-generation smart glasses with AI integration (TED talk).

Posted in AI, News | Tagged agents, audio, hardware, LLM, research, video | Leave a comment

AI Impact Predictions

Posted on 2025-04-18 by KevinYager

Debates about future AI progress or impact are often confused, because different people have very different mental-models for the expected pace, and the time-horizon over which they are projecting.

This figure is my attempt to clarify:

The experimental datapoints come from the METR analysis: Measuring AI Ability to Complete Long Tasks (paper, code/data). The “count the OOMs” and “new regime” curves are extrapolated fits to the data. The other curves are ad-hoc, drawn just to give a sense of how a particular mental model might translate to capability-predictions.

The figure tries to emphasize:

Task complexity covers many orders-of-magnitude. Although imperfect, we can think about the timescale over which “coherent progress” must be made as a proxy for measuring generally useful capabilities.
There are many models for progress, and they vary dramatically in predictions.
Nevertheless, except for scenarios that fundamentally doubt AI progress is possible, the main disagreement among models is over the timescale required to reach a given kind of impact.
The concerns one has (economic, social, existential) will depend on one’s model. (Of course one’s concerns will also be influenced by other assessments, such as the wisdom we expect leaders to exhibit at different stages of rollout.)
It is difficult to define intelligence. Yet, it seems quite defensible to say that we have transitioned from clearly sub-human AI, into a “jagged intelligence” regime where a particular AI system will out-perform humans in some tasks (e.g. rapid knowledge retrieval) but under-perform in other tasks (e.g. visual reasoning). As we move through the jagged frontier, we should expect more and more human capabilities to be replicated in AI, even while some other subset remains unconquered.
The definition of “AGI” is also unclear. Instead of a clear line being crossed, we should expect a greater fraction of people to acknowledge AI as generally-capable, as systems cross through the jagged frontier.

The primary goal of the figure is to clarify discussions. I.e. we should specify which kinds of scenarios we find plausible, which impacts are thus considered possible, and which time-span we are currently discussing.

Posted in AI | Tagged AGI, AI, ASI, impact, RSI | Leave a comment

AI News 2025-04-17

Posted on 2025-04-17 by KevinYager

General

Stanford Human-Centered Artificial Intelligence (HAI) report on the current state of AI: Artificial Intelligence Index Report 2025.

Research Insights

Paper: Reasoning Models Know When They’re Right: Probing Hidden States for Self-Verification. Reasoning generates multiple possible answers, yet this research shows that the correct answer is highly correlated with a correctness signal. This implies that the model could have more efficiently arrived at output this confident/correct answer.
Evaluating the Goal-Directedness of Large Language Models.

LLM

Zyphra releases an open-source reasoning model: ZR1-1.5B (weights, try using).
Anthropic adds to Claude a Research capability, and Google Workspace integration.
OpenAI announces GPT-4.1 models in the API. Optimized for developers (instruction following, coding, diff generation, etc.), 1M context length, etc.; three models (4.1, 4.1-mini, 4.1-nano) provide control of performance vs. cost. Models can handle text, image, and video.
- They also have a prompting guide for 4.1.
- OpenAI have released a new eval for long-context: MRCR.
- OpenAI intends to deprecate GPT-4.5 in the next few months.
OpenAI announces o3 and o4-mini reasoning models.
- These models are explicitly trained to use tools as part of their reasoning process.
- They can reason over images in new ways.
- Improved scores on math and code benchmarks (91-98% on AIME, ~75% on scientific figure reasoning, etc.).
- o3 is strictly better than o1 (higher performance with lower inference cost); o1 will be deprecated.
- OpenAI will be releasing coding agent applications; starting with Codex CLI, which allows one to deploy coding agents easily.
- METR has provided evaluations of capabilities.
- As part of the release, they also provided data showing how scaling RL is yielding predictable improvements.

Safety

OpenAI updates their Preparedness Framework (full document).

Video

Pusa (code, paper) is an open-source video model trained on limited resources.
Alibaba FantasyTalking (paper) can lipsync a character, including body motion (examples).
Tencent Hunyuan announce: Sonic: Shifting Focus to Global Audio Perception in Portrait Animation. Image animation/lipsync that generates more emotive performances (examples).
ByteDance Seaweed-7B: Cost-Effective Training of Video Generation Foundation Model (examples). Can generate audio alongside video, or condition video generation on provided audio.
Kling 2.0 released (launch video). Improved video output, ability to conditionally edit videos, restyle, etc.

Audio

Krisp provides real-time accent modification/neutralization capabilities.

Science

Posted in AI, News | Tagged audio, LLM, research, safety, Science, video | Leave a comment

AI News 2025-04-10

Posted on 2025-04-10 by KevinYager

General

OpenAI announces: ChatGPT Plus is now free for college students through May.
Anthropic Education Report: How University Students Use Claude.
Andrej Karpathy: Power to the people: How LLMs flip the script on technology diffusion.
Mira Murati has raised $2B (at $10B valuation) for her Thinking Machines startup.

Research Insights

New Anthropic results: Reasoning models don’t always say what they think (paper). They find that the plaintext chain-of-though (CoT) of reasoning models may not contain the actual reasoning they used in latent space. This has implications for improving reasoning models, and also suggests (from a safety perspective) that we should not rely on monitoring CoT to infer what models are internally planning.
DeepSeek publish: Inference-Time Scaling for Generalist Reward Modeling. They improve the reinforcement learning step by expending compute to build a better reward model.
Do Larger Language Models Imply Better Reasoning? A Pretraining Scaling Law for Reasoning. While scale helps, one can also encounter overparametrization (memorization instead of reasoning).
DeepResearcher: Scaling Deep Research via Reinforcement Learning in Real-world Environments. They identify behavioral patterns:
- Planning ability emerges naturally in RL, despite not performing SFT on planning data.
- Model verifies answers (even correct answers).
- When retrievals are insufficient, model can generated refined search queries.
- Model can recognize when it lacks sufficient information, and decline to answer.
Rethinking Reflection in Pre-Training. They show that even just from pre-training, models develop some amount of reflective/reasoning understanding.
Concise Reasoning via Reinforcement Learning. They find that RL generically favors longer responses, whereas in reality the correct response is often correlated with being concise. This suggests improving reasoning by favoring shorter answers.
Google Research blog: Geospatial Reasoning: Unlocking insights with generative AI and multiple foundation models. They are building foundation models that integrate the wealth of geospatial data available, which then allows reasoning over this data. Adding more spatial understanding might also unlock other LLM abilities.

LLM

More progress in diffusion language models: Dream 7B: Introducing Dream 7B, the most powerful open diffusion large language model to date.
Meta releases Llama 4 series of MoE LLMs: Scout (109B, 17B active, 16 experts), Maverick (400B, 17B active, 128 experts), and Behemoth (2T, 288B active, 16 experts). These are MoE models with a 10M token context. The models appear to be competitive (nearing the state-of-the-art tradeoff curve for performance/price), and thus extremely impressive for open-source.
- Independent evals (including follow-up) from Artificial Analysis show it performing well against non-reasoning models.
- Evaluation of the 10M context on simple NIAH seem reasonable, but (reportedly) it does not fare as well on deeper understanding of long context.
Cloudflare launch an open beta for their AutoRAG solution.
Nvidia release Llama-3_1-Nemotron-Ultra-253B-v1, which seems to beat Llama 4 despite being based on Llama 3.1.
Amazon announces Nova Sonic speech-to-speech foundation models, for building conversational AI.
Agentica release open-source: DeepCoder-14B-Preview, a reasoning model optimized for coding (code, hf).
Anthropic announce a new “Max” plan for Claude ($100/month).
xAI release an API for Grok-3. Pricing appears relatively expensive (e.g. compared to Gemini models of better performance).
OpenAI adds an evals API, making it easier to programmatically define tests, evaluations, etc. This should make it faster/easier to test different prompts, LLMs, etc.
Bytedance release technical report for Seed-Thinking-v1.5, a 200B reasoning model.
OpenAI add a memory feature to ChatGPT, allowing it to reference all past chats in order to personalize responses.

AI Agents

Cognition AI releases Devin 2.0. Devin has been reframed as an IDE (not unlike Cursor), but they claim that one can use this UI to manage several autonomous software development agents working in parallel.
Detailed review (264 page book): Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems.
Sakana publishes an update to their methods: The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search (code). This system was used to generate a paper that was accepted to a peer-reviewed workshop.
Google’s Deep Research offering now uses Gemini 2.5 Pro, leading to improved responses.
Google unveil Firebase Studio, an AI-assisted coding environment (not unlike Cursor, Bolt, v0) that operates in-browser.
Google announce Agent2Agent (A2A) Protocol, to enhance AI interoperability. It is a complement to MCP.
Google have open-sourced an Agent Development Kit (ADK), a framework for deploying agents.
Google have added Agentspace, a hub for building AI agents.

Audio

Canopy Labs launch Orpheus Multilingual, an open-source text-to-speech (TTS) model.

Image Synthesis

Midjourney unveils their v7 model (currently alpha available to users). It has strong aesthetics (as typical for Midjourney) but prompt adherence and text generation lag behind other models (examples).
HiDream-I1 is a new open-source image model from HiDream.ai (it is now the leading open model).
ByteDance UNO: Less-to-More Generalization: Unlocking More Controllability by In-Context Generation (paper, code, model). Includes generation from input images (e.g. for consistent characters).

Video

Krea introduces Video Re-Style.
ByteDance introduces a robust performance transfer method: DreamActor-M1: Holistic, Expressive and Robust Human Image Animation with Hybrid Guidance (paper).
Runway introduces a turbo version of their newest Gen-4 model.
Paper: One-Minute Video Generation with Test-Time Training (preprint). They add test-time-training (TTT) layers to a pre-trained model, and fine-tune on cartoons. It can generate one-minute video outputs, including shots/cuts that maintain (a semblance of) story consistency. This implies that longer-range video generation (beyond a single clip) can be solved using inference-time compute.

World Synthesis

Microsoft blog: WHAMM! Real-time world modelling of interactive environments. They show generated Quake II gameplay (real-time playable). This shows continued progress towards fully rendered reactive simulations.

Science

Google provide benchmarks for science: Evaluating progress of LLMs on scientific problem-solving.
A researcher at Brookhaven National Laboratory (Weiguo Yin) used o3-mini-high to find new solutions to a physical model: Exact Solution of the Frustrated Potts Model with Next-Nearest-Neighbor Interactions in One Dimension: An AI-Aided Discovery.
Coordinated AI agents for advancing healthcare (pdf).
Training state-of-the-art pathology foundation models with orders of magnitude less data.
Google report on AI for medicine:
- Towards conversational diagnostic artificial intelligence
- Towards accurate differential diagnosis with large language models

Brain

Full mapping of a cubic millimeter of mouse brain: Biggest brain map ever details huge number of neurons and their activity (scientific paper: Functional connectomics reveals general wiring rule in mouse visual cortex).

Hardware

Google unveils their next-generation tensor processing chip (TPU): Ironwood: The first Google TPU for the age of inference. Each chip can deliver 4,614 TFLOPs.
- Reportedly, Ilya Sutskever’s startup, Safe Superintelligence (SSI), is using Google’s TPUs.

Robots

Some short clips of 1X Neo performing domestic tasks autonomously.
- NY Times: Invasion of the Home Humanoid Robots. Dozens of companies are building robots that look like humans. One of them is training a machine to be a butler and will soon test them in homes.
Westwood Robotics is working on Themis v2.
Clone show an update to their Protoclone humanoid; >1,000 artificial muscles controlling >200 degrees-of-freedom.

Posted in AI, News | Tagged agents, audio, brain, hardware, image, LLM, multimodal, research, robots, Science, video, world synthesis | Leave a comment

AI News 2025-04-03

Posted on 2025-04-03 by KevinYager

General

AI timeline visualization: The Road to AGI (2015–2025), by James Campbell and Emiliano Garcia-Lopez (code).
Anthropic releases the second report for their economic index: Anthropic Economic Index: Insights from Claude 3.7 Sonnet.
xAI buys the 𝕏 (formerly Twitter) social media platform. The internal all-stock deal values xAI at $80B and 𝕏 at $33B ($45B minus $12B debt).
OpenAI raises an additional $40B at $300B valuation.
LLMs shown to be viable for therapy: First Therapy Chatbot Trial Yields Mental Health Benefits.
- Paper: Randomized Trial of a Generative AI Chatbot for Mental Health Treatment.
LLMs as viable tutors: LLM Support for Tutors GPT-4 boosts remote tutors’ performance in real time, study finds.
- Paper: Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise.
New study evaluates human ability to detect AI in a more stringent setting: Large Language Models Pass the Turing Test.
- Note earlier work already showing LLMs passing less-stringent Turing Tests:
  - 2023-05: Human or Not? A Gamified Approach to the Turing Test
  - 2023-10: Does GPT-4 pass the Turing test?
  - 2024-05: People cannot distinguish GPT-4 from a human in a Turing test
  - 2024-07: GPT-4 is judged more human than humans in displaced and inverted Turing tests

Research Insights

Meta preprint: Multi-Token Attention. They combine attention (query, key, head operations) over multiple tokens; convolution operations allow nearby queries/keys to affect each other’ss attention weights.
Danijar Hafner et al. (Google DeepMind) present DreamerV3.
- Media report: AI masters Minecraft: DeepMind program finds diamonds without being taught. The Dreamer system reached the milestone by ‘imagining’ the future impact of possible decisions.
- Paper: Mastering diverse control tasks through world models.

Safety

Control AI releases: The Direct Institutional Plan. They suggest designing policies that prevent development of superintelligence, and spreading awareness among democratic institutions.
Google DeepMind: DeepMind: Taking a responsible path to AGI.
- Paper: An Approach to Technical AGI Safety and Security

LLM

OpenAI pushed an update to their 4o model. This has significantly improved its ranking (e.g. now best non-reasoning model on coding benchmark).
An interesting test of GPT-4o in-context image generation: it is unable to generate an image of a maze with a valid solution; at lest when the maze is a square. However, if you ask it to make an image of a diamond orientation maze (45° rotated square), it succeeds to have a valid solution. We can rationalize this based on the sequential order of autoregressive generation. By generating first from the start of the maze (and only its local neighborhood), and similarly finishing with this sort of locality, the model can more correctly build a valid solution. (Conversely, the usual square orientation requires longer-range reasoning across image patches.)
- At first, this might seem like just another silly oddity. But it shows how recasting a problem, just by changing the generation order, can massively change model performance. This sheds light on how they “think” and suggests that alternate generation strategies could perhaps unlock capabilities.
  - For instance, one could imagine an LLM with different branches (like MoE?) where each branch is trained on a different autoregression strategy (left-to-right, right-to-left, block diffusion, random, etc.) such that the overall LLM can invoke/combine different kinds of thinking modes.
- Another trick is to ask it to generate an image of a maze with the solution identified, and then update the image to remove the solution. This is a visual analog of “think step-by-step” and other inference-time-compute strategies. This implies that current models have untapped visual reasoning capabilities that could be unlocked by allowing them to visually iterate on problems.
Anthropic announces Claude for Education, which provides a university-wide solution tailored to education.

AI Agents

Amazon introduce Nova Act, a research preview for agents controlling web browsers.
AI Digest has started an experiment: they launched 4 computer-use agents, and gave them the task of getting donations for a charity of their choice. The agents can chat to each other, and human visitors can also chat with them. They have begun to (slowly) work on the problem. You can view their ongoing activities here.
General Agents claims they have a general-purpose computer-use agent (Ace) that operates your local computer.
OpenAI release a new benchmark: PaperBench: Evaluating AI’s Ability to Replicate AI Research (paper, code).
Zapier adds MCP support, so AI agents can now access a very broad range of web apps (Slack, Google Sheets, Notion, etc.).

Audio

ElevenLabs:
- Adds native, low-latency RAG for conversational AI.
- Launch Actor Mode, where you can use your voice to direct the AI’s performance.
Udio introduces Styles, allowing generation from a provided audio clip.
Mureka AI enables more fine-grained music generation.

Image Synthesis

Ideogram 3.0 released.
OpenAI adds regional selection to their new in-context 4o image generator, allowing tailored updates to images.

Video

Runway ML launches a new model: Gen-4. Improvements in realism, physics, character consistency, etc. Sample short videos: The Lonely Little Flame, The Herd, The Retrieval, NYC is a Zoo, Scimmia Vede. (More examples.)
Meta unveils: MoCha: Towards Movie-Grade Talking Character Synthesis (preprint). Remarkably good human character generation from audio input.
Higgsfield is showing good camera control.
Sync introduces lipsync-2, which generates expressive synced video that maintains expressiveness.

Science

Curated list of science datasets: Awesome Materials & Chemistry Datasets.

Robot

KEENON Robotics introduces wheeled humanoid XMAN-R1.
Unitree Dex5 Dexterous Hand (20 degrees of freedom).
More video of Figure robots performing real work in a BMW factory.

Posted in AI | Tagged agents, audio, image synthesis, LLM, robot, robots, safety, Science, video | Leave a comment

AI News 2025-03-27

Posted on 2025-03-27 by KevinYager

General

This week saw a remarkable amount of advancements and releases, which is fitting since it is the two-year anniversary of “the biggest week in AI history” (March 2023: GPT-4 release, AI copilot for Microsoft 365 announced, Bing Image Creator preview, Github Copilot X announced, Midjourney v5 released, Bard early access, Google Docs generative tools announced, Anthropic releases Claude chatbot, high-efficiency Alpaca model released open source, Runway ML teases Gen2 text-to-video product, “Sparks of AGI” paper, etc.).
- (This level of activity is now considered a normal week.)
Perplexity AI in talks to double valuation to $18 billion, raise up to $1 billion in new funding.
New study: The Cybernetic Teammate: A Field Experiment on Generative AI Reshaping Teamwork and Expertise. Ethan Mollick blog post: The Cybernetic Teammate. Having an AI on your team can increase performance, provide expertise, and improve your experience.
Epoch AI:
- Most AI value will come from broad automation, not from R&D.
- GATE: Modeling the Trajectory of AI and Automation (docs, paper). Their economic model suggests large investments in AI, with subsequent very large productivity gains.
OpenAI and MIT Media Lab: Early methods for studying affective use and emotional well-being on ChatGPT.
- Investigating Affective Use and Emotional Well-being on ChatGPT.
- How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use: A Longitudinal Controlled Study.
- These studies show a diversity of outcomes. Overall, the outcome (positive or negative) depends on how it is used.
ARC Prize officially release the ARG-AGI-2 competition (2025). Machine Learning Street Talk partnered to produce a launch video.

Research Insights

Compute Optimal Scaling of Skills: Knowledge vs Reasoning. They find different scaling behaviors for knowledge vs. reasoning improvements.
Google report on analogies between artificial neural nets and human brains: Deciphering language processing in the human brain through LLM representations. They find that neural activity in human brains aligns cleanly with LLM embeddings.
Interruption is All You Need: Improving Reasoning Model Refusal Rates through measuring Parallel Reasoning Diversity. They reduce hallucinations by inserting interruption tokens and sampling in parallel.
Unlocking Efficient Long-to-Short LLM Reasoning with Model Merging (code). They show that model merging can more efficiently leverage the combination of system 1 and system 2 behaviors.
I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders. There are feature vector directions corresponding to reasoning; one can increase or decrease reasoning effort by adjusting this vector.
Anthropic: Tracing the thoughts of a large language model. They investigate neural circuits in Claude.
- Paper: Circuit Tracing: Revealing Computational Graphs in Language Models. They replace the typical computational steps with more interpretable analogs, making it easier to track the circuits.
- Paper: On the Biology of a Large Language Model.

LLM

Anthropic adds web search to Claude.
Modifying Large Language Model Post-Training for Diverse Creative Writing.
DeepSeek release DeepSeek V3-0324 (685B); highest-scoring non-reasoning model.
Google unveil Gemini 2.5 Pro, an LLM with state-of-the-art performance on many benchmarks (including #1 on lmarena, and record-setting 18% on Humanity’s Last Exam) and 1M context length (with good performance at large context).

Multimodal

Nvidia introduce the Nemotron-H family of models (8B, 47B, 56B), including base/instruct/VLM variants. They are hybrid Mamba-Transformer models that achieve good efficiency.
Cohere introduce Aya Vision, open-weights vision models (8B, 32B).
Alibaba release Qwen2.5-VL-32B-Instruct (hf), with RL optimization of math and problem-solving abilities (including for vision).
MoshiVis is an open-source vision/speech model (try here), that can converse naturally about images.
OpenVLThinker: An Early Exploration to Complex Vision-Language Reasoning via Iterative Self-Improvement (code).
Alibaba release Qwen2.5-Omni-7B (tech report, code, weights, try it). Multimodal: text, vision, audio, speech.

AI Agents

OpenAI adds support for Anthropic’s Model Context Protocol (MCP), solidifying it as the standard mechanism for giving AI agents access to diverse resources in a uniform way.
- List of MCP servers
- Registry of MCP servers

Safety

Superalignment with Dynamic Human Values. They treat alignment as a dynamic problem, where human values may change over time. The proposed solution involves an AI that breaks tasks into smaller components, that are easier for humans to guide. This framework assumes that alignment of sub-tasks correctly generalizes to desirable outcomes for the overall task.
Google DeepMind: Defeating Prompt Injections by Design.

Audio

OpenAI announced new audio models: new text-to-speech models (test here) where one can instruct it about how to speak; and gpt-4o-transcribe with lower error rate than Whisper (including a mini variant than is half the cost of Whisper).
OpenAI update their advanced voice mode, making it better at not interrupting the user.

Image Synthesis

Tokenize Image as a Set (code). Interesting approach to use an unordered bag of tokens (rather than a serialization, as done with text) to represent images.
StarVector is a generative model for converting text or images to SVG code.
Applying mechanistic interpretability to image synthesis models can offer enhanced control: Unboxing SDXL Turbo: How Sparse Autoencoders Unlock the Inner Workings of Text-to-Image Models (preprint, examples).
The era of in-context and/or autoregressive image generation is upon us. In-context generation means the LLM can directly understand and edit photos (colorize, restyle, make changes, remove watermarks, etc.). Serial autoregressive approaches also handle text and prescribed layout much better, and often have improved prompt adherence.
- Last week, Google unveiled Gemini 2.0 Flash Experimental image generation (available in Google AI Studio).
- Reve Image reveal that the mysterious high-scoring “halfmoon” is their image model, apparently exploiting some kind of “logic” (auto-regressive model? inference-time compute?) to improve output.
- OpenAI release their new image model: 4o image generation. It can generate highly coherent text in images, and iterate upon images in-context.
  - This led to a one-day Ghibli-themed spontaneous meme explosion.
  - It is interesting to see how it handles generating a map with walking directions. There are mistakes. But the quality is remarkable. The map itself is mostly just memorization, but the roughly-correct walking directions and time estimation point towards a more generalized underlying understanding.

Video

SkyReels is offering AI tools to cover the entire workflow (script, video, editing).
Pika is testing a new feature that allows one to edit existing video (e.g. animating an object).

World Synthesis

Wayve reports: GAIA-2: Pushing the Boundaries of Video Generative Models for Safer Assisted and Automated Driving (technical report).

Science

Vant AI announces Neo-1, an atomistic foundation model useful for biochemistry and nanoscience.
Google release TxGemma models (2B, 9B, 27B): Open models to improve therapeutics development.

Hardware

Halliday: smart glasses intended for AI integration ($430)

Robots

Unitree shows a video of smooth athletic movement.
Figure reports on using reinforcement learning in simulation to greatly improve the walking of their humanoid robot, providing it with a better (faster, more efficient, more humanlike) gait.
Google DeepMind paper: Gemini Robotics: Bringing AI into the Physical World. They present a vision-language-action model capable of directly controlling robots.

Posted in AI, News | Tagged agents, audio, hardware, image, image synthesis, LLM, multimodal, research, robots, safety, Science, video, world synthesis | Leave a comment

AI News 2025-03-20

Posted on 2025-03-20 by KevinYager

General

Kevin Roose, New York Times: Powerful A.I. Is Coming. We’re Not Ready.
Predictions: Nicholas Carlini: My Thoughts on the Future of “AI”.
- “I have very wide error bars on the potential future of large language models, and I think you should too.”
Predictions: Glimpses of AI Progress: Mental models for fast times.
Strategic opinion piece: AI Dominance Requires Interpretability and Standards for Transparency and Security.

Research Insights

Google DeepMind: Mixture-of-Depths: Dynamically allocating compute in transformer-based language models.
Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models (project, code, hf). They combine autoregressive and diffusion approaches to text generation.
Research from METR: Measuring AI Ability to Complete Long Tasks. A very valuable way to gauge AI utility is to compare to the length of the equivalent human effort for the task. As AI improves in coherence, we can expect it to tackle progressively longer-horizon tasks.
- Media report (Nature): AI could soon tackle projects that take humans weeks.

LLM

Baidu announce Ernie 4.5 and X1 (use here). They claim that Ernie 4.5 is comparable to GPT-4o, and that X1 is comparable to DeepSeek R1; but with lower API costs (Earnie 4.5 is 1/4 the price of 4o, while X1 is 1/2 of R1). They plan to open-source the models on June 30th.
Mistral release Mistral Small 3.1 24B. They report good performance for the model size (e.g. outperforming GPT-4o-mini and Gemma 3).
LG AI Research announce EXAONE Deep, a reasoning LLM (2.4B, 7.8B, 32B variants; weights) that scores well on math benchmarks.
Nvidia release Llama-Nemotron models, which can do reasoning (try it here).

Safety

Anthropic: Auditing language models for hidden objectives (paper).
Reducing LLM deception at scale with self-other overlap fine-tuning.
- See earlier post: Self-Other Overlap: A Neglected Approach to AI Alignment.
- And paper: Towards Safe and Honest AI Agents with Neural Self-Other Overlap.
- The method essentially enforces a form of empathy, by training the AI to align its self-representation with the representation of others.
Joe Carlsmith continues in his series: How do we solve the alignment problem? The latest essay is: AI for AI safety (audio version).
Fei-Fei Li et al. have drafted a report: Joint California Policy Working Group on AI Frontier Models.

Vision

Thera: Aliasing-Free Arbitrary-Scale Super-Resolution with Neural Heat Fields (code, use).
SmolDocling: An ultra-compact vision-language model for end-to-end multi-modal document conversion (hf). A 256M vision-language model optimized for converting document images into structured text.
Vamba: Understanding Hour-Long Videos with Hybrid Mamba-Transformers (preprint, code, hf).

Image Synthesis

Gemini 2.0 Flash Experimental (available in Google AI Studio) is multimodal, with image generation capabilities. By having the image generation “within the model” (rather than as an external tool), one can iterate on image generation much more naturally. This incidentally obviates the need for more specialized image tools (can do colorization, combine specified people/places/products, remove watermarks, etc.).

Video

ReCamMaster: Camera-Controlled Generative Rendering from A Single Video (preprint, code) Allows one to modify the camera position/motion on an existing video clip (try it here).
WarpTuber enables real-time VTubing with AI avatar models (tutorial video).
Sync labs tease a lip-sync model.
Stability AI announces an image-to-video model that generates camera motion over the estimated 3D scene (code).

Audio

Udio have distilled their v1.5 model into a faster and higher-quality variant: Allegro.
Canopy Labs Orpheus 3B is a high-quality open-source text-to-speech model (weights: pretrain, finetune; try using).
AudioX: Diffusion Transformer for Anything-to-Audio Generation.

Science

Med-R1: Reinforcement Learning for Generalizable Medical Reasoning in Vision-Language Models.

Robots

Figure reports: BotQ: A High-Volume Manufacturing Facility for Humanoid Robots. Initially targeting 12k robots/year; eventually ramping up to 100k/year.
EngineAI video shows fluid dancing (behind the scenes: teaching, filming).
Nvidia (in partnership with Google DeepMind and Disney Research) announces announces Newton, an open-source physics engine for robotic simulation.
Boston Dynamics video of Atlas showing improved athletics. Of note is that this is accomplished using reinforcement learning in simulations.

Posted in AI, News | Tagged audio, image synthesis, LLM, research, robots, safety, video, vision | Leave a comment