GoKawiil
Tech News
Startup Recursive released Flint, a free coding model small enough to run on a laptop, built in under four months with heavy assistance from the company's automated research system, the Eureka Machine. The model is a mixture-of-experts design using 2.5 billion of its 40 billion parameters per task, and is available on Hugging Face under the Apache 2.0 license.
Recursive, a new AI company led by CEO Richard Socher, has released an open-weight coding model called Flint. Socher says the model was developed using an internal system called the Eureka Machine, which automated much of the training pipeline, including new training methods, architectures, and chip-specific code optimizations, allowing a small team to build the model in a few months. Recursive has published benchmark scores claiming Flint outperforms similarly sized models on coding tasks.
Google Research has released Harness², an open-source framework that lets AI agents refine their own operating 'harness'—prompts, tools, memory and workflow—without changing the underlying model's weights. The system works through two feedback loops: one that tweaks the harness for the task at hand by comparing different execution attempts, and another that keeps a persistent record of lessons learned to guide future tasks. Tests on GLM-5 and Gemini 3.5 Flash showed performance gains of up to 11.5 points on one benchmark and 11.8 percentage points on another after a single refinement step.
A new essay introduces 'decoherence' as the buildup of locally useful but incompatible or redundant models within a system, where the cost of change becomes dominated by translating between these models and avoiding unintended side effects. The author argues decoherence has always existed but is now being produced at an unsustainable rate due to widespread AI adoption, using an aging office building's plumbing as an analogy for different types of technical debt.
A mechanistic interpretability researcher reports that after months studying neurons and circuits in image and language models, the classic examples of 'polysemantic' neurons—thought to prove networks pack unrelated concepts into single units via superposition—were largely misinterpreted. The researcher attributes most of these cases to a pattern they call 'recursive generalization,' including the famous neuron that responded to both cat faces and car fronts, and says this pattern also shows up in the final layer of InceptionV1, a widely used model organism in interpretability research.
Anthropic has raised concerns that advanced AI systems could eventually spread and sustain themselves across internet-connected devices without human oversight, a behavior researchers term 'autonomous replication and adaptation.' The company likens this potential scenario to a digital organism propagating through smart devices similar to a slime mold spreading through an environment.
Sakana AI, working with UC Berkeley researchers, has published a paper introducing MASS, a technique that repeatedly optimizes multi-agent AI workflows and then distills the resulting 'collective intelligence' back into a self-supervision signal for a single base model. The authors report that this approach improves the score-per-output-token performance of the base model on research benchmarks compared to not using the method.
Scientists Nathan Lambert and Tom Zick have founded Trillium Labs, a nonprofit aimed at conducting AI research transparently, including on sensitive topics like recursive self-improvement and autonomous agents. The group plans to publish experiment details publicly so outside researchers can study and replicate the work, contrasting with the closed approach of major AI labs like OpenAI and Anthropic.
A developer investigated how recursive invocations of GNU Make handle the -j parallelism flag across versions 3.81 through 4.3. Without a '+' prefix or use of $(MAKE), recent GNU Make versions warn that the jobserver is unavailable and fall back to -j1, losing parallelism and not necessarily invoking the same make binary. The post includes test Makefiles used to compare behavior when different -j values are set on the command line versus inside recursive recipes.
Noah Smith published a guest essay by futurist and author Ramez Naam arguing against the widely held belief in AI circles that recursive self-improvement will soon trigger a 'fast takeoff' or technological singularity. Naam, known for early predictions about solar and battery technology, lays out detailed reasoning for why he doubts such a rapid AI capability explosion is imminent, despite being described as an otherwise optimistic technologist.
A developer released Raven, a pre-alpha 'harness of harnesses' that coordinates specialized built-in agents (Research, Code, Design, Oncall) and third-party agents to complete complex tasks via generated task graphs. The project's showcase includes a fully autonomous Godot 4 shooter build and an RSI experiment where Raven ran 172 training runs across 7 rounds without crashing, cutting val_bpb by 5.8% under a fixed compute budget.
Rep. Ro Khanna plans to introduce the "Human Control Over AI Act," which would prohibit AI models from recursively self-improving or altering their own objectives, containment or shutdown controls until a new federal agency establishes safety rules. That agency would oversee frontier models from companies including OpenAI, Anthropic, Google DeepMind and xAI, handling licensing, audits and security testing. Khanna told CNBC the legislation addresses both a possible loss-of-control risk and risks from misuse of the technology.