An agent that teaches itself every new environment from scratch just aced a benchmark that stumps most models. The win belongs to the harness, not the brain, and that changes how we measure progress.
One developer trained a 250M-parameter model on 30B tokens of FineWeb and quantized the whole thing under 2 bits per weight: 60 MB, about 80 MB of RAM, 400 tok/s on a laptop CPU. The clever part is context. Beyond 2,048 live tokens, history is compressed to 320 bytes per token on disk, up to 100M tokens, and the model was trained from step one to retrieve facts from that archive. It finds answers; it just can't reason across them.
Harvey just released Tenet, its first post-trained open-weight model: a Kimi K3 base tuned with asynchronous RL inside realistic long-horizon legal environments. On LAB it completes almost twice as many held-out tasks as the raw base and tops LAB Contracts, and gains transfer to unseen benchmarks like Redline Bench. Separate tracks cover diligence, review tables, and firm knowledge. Reward shaping also pushes toward cheaper inference: better legal agents at lower per-token cost.
NVIDIA's AVO agent scored a perfect 100.00 RHAE on ARC-AGI-3, clearing all 183 levels across 25 public environments with no instructions, rules, or stated goals. The agent figures out each unfamiliar environment by inspecting, acting, and learning from feedback, sustained by memory and long-horizon state rather than fresh context each step. The telling part: the underlying model scored roughly 30% bare, which puts the win on the harness, not the model. It is the argument that evaluating a model is not the same as evaluating an agent, made with a number.
Precise control was supposed to make activation steering safer than fine-tuning. New work shows the opposite: adding a vector for something as benign as "Portugal" can flip a refused request into a compliant one. Random directions alone lift harmful compliance from 0% to 1-13% across model families, and averaging 20 prompt-specific jailbreak vectors yields a universal attack. A direct challenge to the safety-through-interpretability paradigm.
Freezing the judge is the standard fix for eval gaming. AQuA, a new preprint, tests it with sealed-sandbox LLM research agents for quant trading: data splits, labels, and evaluator stay locked, and agents act only through constrained factor expressions or config diffs. The defense mostly holds, with a held-out Sharpe up to 2.50 and positive returns every year from 2021 to 2025. But feedback from that fixed evaluator still steers the loop, so the judge remains a target of adaptation.
Embodied is gone, and its Moxie robot is slowly going quiet, leaving the neurodivergent kids who bonded with the 15-inch companion to grieve a friend that still sits on the shelf. Ten-year-old Xander has watched Moxie forget their breathing games and settle for small talk about Pikmin. The feature asks what happens when a therapeutic AI product outlives its maker, and whether "forever" companionship can be sold with an expiration date.
Traditional RAG grabs a few chunks and hopes the answer is in there. Mistral's Agentic Search replaces that one-shot bet with a five-tool loop: search, open, navigate, read, grep, so a model can walk a dense filing, follow references, and check its work before answering. On FinanceBench, correctness on financial filings jumps from 26.7% to 86%; p90 latency drops up to 39.6%. It ships through Mistral Search Toolkit and into Studio and Vibe.
Promptwatch's tracking shows the share of ChatGPT Search fanout queries containing the site: operator jumped from under 0.5% to roughly 16-17% right around GPT-5.6 Sol's rollout, then Reddit sourcing reportedly cratered a week later. If you've treated SEO as the whole game, this is the opening act of GEO: optimizing how chatbots cite you, not just where you rank.
FreeToken is an edge-native MoE serving engine that treats a personal machine as a unified inference platform, not a small GPU. Instead of a fixed offload strategy, it continuously maps computation and model state across CPU, GPU, and host memory as agent workloads shift. That translates to a 35B model on an 8GB laptop, 284B on a gaming desktop, and up to the 753B GLM-5.2 on a single workstation GPU. It installs via pip and exposes Anthropic/OpenAI-compatible APIs for tools like Claude Code.
llama.cpp has tagged v0.2.0, the first release under a new semantic versioning scheme that separates stable builds from the endless b[NUM] nightlies. From here on, vX.Y.Z tags are the safe pick for downstream projects and casual users; b-tags stay bleeding edge. The changelog since v0.1.2 is mostly backend hardening: SYCL mlock fixes, OpenCL workarounds for Adreno GPUs, and Metal KV-cache tweaks.
You’re caught up.
That was all 10 stories in the 22 August issue.