New Anthropic research: Persona vectors: Monitoring and controlling character traits in language models. Many interesting results, including inducing a particular behavior by adjusting activations in a particular direction. Used at inference-time, this can induce or inhibit behavior, but at a cost in capability (as previously known). E.g. one can steer away from the “evil” direction, but one worsens model task performance. But interestingly, one can steer the model during training to prevent certain behaviors from ever being learned. Counter-intuitively, one actually steers towards an undesired behavior (e.g. in the evil direction) during training. This acts as a sort of inoculation, since the model doesn’t need to add the over-emphasized behavior to its learned weights; and at runtime (when the bias is no longer present) it snaps back to desired behavior (e.g. towards the good direction).
OpenAI announces the release of two open-weight reasoning models: gpt-oss-120b (for servers or high-end desktops) and gpt-oss-20b (for desktop/laptop). Local reasoning model (full access to chain-of-thought) that should be good for agentics (hf, github, test). Supposedly similar in capability to o4-mini.
Reasoning model that selects the right amount of compute. Multiple models behind the scenes: GPT5 (default), GPT5-mini, GPT5-nano, GPT5-Pro (for Pro tier only). Available in API.
It’s better. Strong performance across many metrics: 75% on SWE-bench, 84% MMMU, 100% AIME 2025. Better writing, better coding. Improved voice. Can see via video input.
Can now select among different “personalities”.
Trained (in part) by using o3 to generate teaching datasets.
Google paper provides insight into how LLMs can do in-context adaptation: Learning without training: The implicit dynamics of in-context learning. They find that the transformer architecture leads to each token generating a “temporary patch” that steers the model, as if it were fine-tuned on the context data.
New paper with a bold claim: AlphaGo Moment for Model Architecture Discovery. They claim that searches through different AI/Ml architectures can consistently yield discoveries/improvements, including surprises humans might not have considered.
LLM
Google NotebookLM rolls out a new user interface, and adds video overviews.
Memories.aiintroduces a Large Visual Memory Model. The claim is that it enables reasoning over video with nearly unlimited context length. This is accomplished by generating concise semantic codes for each frame, allowing search over long videos.
Hunyuan announceHunyuan3D World Model 1.0 (github, huggingface). Currently supports movement within a small circle from the initial viewpoint. But helps point the way towards more general-purpose sims.
In general, tech adoption is faster now than in the past. AI adoption is even faster than the modern typical trend
There are interesting parallels between human psychology and AI responses. New research: Call Me A Jerk: Persuading AI to Comply with Objectionable Requests. Shows that LLMs can be persuaded using methods similar to humans. This doesn’t mean LLM internal cognition is humanlike, but at least shows how LLMs are reproducing (modeling) human foibles.
Learning from one and only one shot. A new “information lattice learning” approach shows good generalization (better than transformer, for certain visual tasks).
OpenAI goldwas done with an LLM not using tools (theorem prover, etc.). It was a special unreleased LLM; but it was not optimized on math problems per se, but rather implementing some new kind of improved algorithm (presumably RL?). This level of capability is expected to be included in a future OpenAI model (but not GPT-5).
There is also a claim of using o4-mini-high to achieve IMO gold (using the Crux agent framework).
Taken together, these results suggest that reasoning models are already at a very high level of math competence. Moreover, it adds some evidence that existing LLMs have many latent skills, that RL can unlock.
METR analysis: Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (blog, commentary/analysis). Their results show that use of AI tools by programmers decreased productivity. This is surprising to the large number of programmers who use LLMs to improve their work. One caveat is that this was analyzing productivity gains for developers working on their own codebases (whereas the biggest gains for LLM usage comes from exploring unfamiliar topics/code). One might also wonder how the results change for free-form use instead of as prescribed in this study. Nevertheless, this does weaken the argument for AI already delivering productivity gains.
METR is showing results for a wider range of tasks (this appears to be an update to their earlier report from May 2025). The landmark METR result showed that AI execution of software engineering tasks (as measured by “task time” for a human to complete) was increasing exponentially. Now they add a variety of other tasks to the evaluation. Each tasks shows some form of exponential scaling, though the magnitude and scaling differ.
It is worth remembering that there will be tradeoffs between performance and monitorability. Various “thinking in latent space” approaches have demonstrated improved performance, but that obfuscates AI thinking (requiring imperfect mechanistic interpretability to then attempt to recover internal processes). We should be willing to give up some immediate performance gains in order to increase our ability to align (which will yield more long-term gains).
LLM
Kimi AIreleasesKimi-K2, an open-source (permissive license) 1T parameter model (MoE, 32B active). It performs extremely well across multiple benchmarks, especially coding, despite being a non-reasoning model (try it, API, code, weights).
OpenAI launches Agents (video). It uses a combination of text-browsing agentics (like Deep Research) and visual-browsing (like Operator) to handle open-ended asynchronous tasks. It can access connectors (Google Drive, etc.), tools (image generation), and can generate (e.g.) slide decks. Achieves 42% on Humanity’s Last Exam and 27% on FrontierMath.
Runway is starting to deploy Act-Two, an improved motion capture model that can transfer a video performance to an AI avatar based on a single input image.
Predicting thinking time in Reasoning models. By predicting estimate token needs during CoT reasoning, feedback to user (e.g. progress bar) could be provided. Models internally have state that is predictive of further token usage.
gremllm is clever and/or diabolical. It is a Python library that generates on-the-fly the attributes and methods of a Python object. Thus, one need not actually define the methods for a new class; simply allow the LLM to hallucinate them when they are called for.
Although this sounds silly and dangerous, there are viable use-cases. In March 2023 (site and code no longer online), there was some exploration of “imaginary programming” wherein one would define a function’s requirements but never actually code the function (the LLM would instead stand-in for the function at call time).
xAI release Grok 4 (and Grok 4 Heavy). Benchmarks are strong, taking the lead on several, including 100% on AIME, 44% on Humanity’s Last Exam, and 16% on ARC-AGI-2 (c.f. 9% Claude Opus 4). If real-world utility matches benchmarks, then Grok 4 may take the lead as the best model.
Now (2025-06): Prediction of LLMs being scaled down into “cognitive cores”; small edge-optimized (on-device inference) LLMs that have minimal knowledge but maximized reasoning and tool-use abilities. Can rapidly iterate to retrieve required results and build answers.
The Automated LLM Speedrunning Benchmark: Reproducing NanoGPT Improvements. The benchmark measures agentic abilities by asking the agent to improve the training speed for a small LLM (as a proxy for more general “AI recursive self-improvement”). Current agents do surprisingly badly (sub-human performance even with significant hints). Going forward, this eval (or variants thereof) should prove useful to measure “useful agentic” performance.
LLM
Inception Labslaunch Mercury, a diffusion LLM. The fast inference of diffusion architecture puts it in a new regime for speed-vs-performance.
Tesla shows off a car self-driving (no occupants) from factory to customer’s address. No doubt the route was carefully selected and vetted. Nevertheless, it is impressive.
Tesla launched a limited rollout of their full-self driving Robotaxi (with in-vehicle employee monitor, for now), in Texas.
Robots
K-Scaleannounces that you can now order one of their open-source humanoid robots (9k$ early adopter price; 16k$ nominal price).
Dense SAE Latents Are Features, Not Bugs. They find pairs of opposing features that fire very frequently. Far from being useless, they find these encode meaningful concepts.
Sakana AI: Reinforcement Learning Teachers of Test Time Scaling (preprint). Rather than using to RL to improve solution, they focus on improving the ability for the model to teach other models. RL reward is based on how useful generated training examples are (to smaller learning models); rather than being rewarded on correctness of their final answer.
LLM
Google releasesGemma 3n, for on-device/edge multimodal processing.
Distillation Robustifies Unlearning (preprint, demo, discussion). Normal unlearning suppresses knowledge in a model, but adversarial prompting or fine-tuning can bring the knowledge/behavior back. They show that distilling into a new model more reliably eradicates the undesired information.
They show that intentional misalignment training (e.g. to write bad code) causes an emergent “evil” personality. But this can be detected and countered.
AMD unveiled its new MI350 chip, optimized for AI workloads. They are focusing on open/compliant coding standards, and energy/cost efficiency.
Cars
Data from Waymo: New Insights for Scaling Laws in Autonomous Driving. They show scaling laws apply to autonomous driving: using more data and more compute for training yields reliable improvements in performance.
Robots
RoboBrain 2.0 is an open-source, general purpose robot control model (video).
Epoch AI was given access to the full reasoning traces of o3-mini (normally only summaries are shown to the user) to conduct this research: Beyond benchmark scores: Analyzing o3-mini’s mathematical reasoning. Mathematicians reviewed the traces of working on math problems; one evaluated described o3 as a “vibes-based inductive reasoner”.
An interesting result from the report (section 7.2: Eating the multimodal free lunch): They base model is multi-modal, but RL is done using text only. Yet, they observe this text-only training does not harm multi-modal performance; in fact multi-modal performance improves. This suggests modalities are well-entangled and that transfer learning between modalities is naturally occurring.
4DV is demoing 4D Gaussian Splatting, wherein multi-camera video data is converted into a temporal/video 3D-spatial reconstruction (videos showing interaction: 1, 2, 3).
Tesla has provided some updated details on their current “full self-driving” (FSD) implementation. Some claims: 3.5B miles driven by FSD across 6 million vehicles, 54% safer than human.
The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (code). Interesting result showing that one can improve model performance by training against negative signals (failure). In fact, this has certain advantages; e.g. purely rewarding successes means that alternative success pathways are not reinforced, whereas negative signals to failure boosts all success pathways.
General agents need world models. They provide formal evidence that any agentic system that achieves generalized capability must be building and exploiting some kind of world model.
LisanBench (github) is a new benchmark that evaluates long-term task coherence (“stamina”) through a game where the LLM must progressively alter a word (one character at a time), always yielding a valid English word, to build the longest possible chain. Although highly contrived, this does seem to test longer-range planning. The results conform to vibes about model intelligence.
Anthropic has launched Claude Explains, a blog of AI generated posts (with human verification). The focus (currently) appears to be teaching simple coding concepts.
Yoshua Bengio launchesLawZero, a non-profit dedicated to advancing safe-by-design AI.
Audio
Elevenlabs introduces a multi-modal assistant, that can handle mixture between voice and text input (at the same time; not requiring toggling between modes). It does seem like a productive way to interact with an AI.
Blandannounces an improvement in their text-to-speech model, with cloning of voice, accent, style, etc. They claim it is finally past the uncanny valley.
