Built on a larger base model with a longer RL run and an entirely new safeguard stack, Grok 4.7 hits CursorBench 4.0 frontier price-performance at $2/$6 per million tokens; Nous Portal ran it at half price for a week.
Assuming the model will be tricked, Muse never hands it real credentials, runs tools in isolated Linux containers, and relies on an independent gatekeeper to verify outbound calls.
Overseas sales of its AI accelerators are paused while China's own demand for 15,488-chip Atlas clusters outstrips supply; optical networking scales clusters toward 120 EFLOPS to offset a per-chip gap versus Nvidia.
Because an agent's harness (prompts, control flow, tooling, memory) drives capability, this work regularizes recursive self-improvement of the harness itself, curbing memorization of training tasks while preserving gains.
A differentiable, fixed-size routed-memory KV cache learns which information to retain in long-context inference, beating selective eviction and pruning baselines by keeping core context in a learned sparse control path.
The real limiter at extreme compression rates is how the budget is spread across heads; co-optimizing per-head low-rank truncation and scalar quantization dominates uniform allocation, with the biggest gains at low bit-rates.
Agents evolve their own capabilities online in continuously interactive environments, treating environmental knowledge as living input rather than fixed data, and beat RL baselines that adapt to fixed task distributions.
Training LLMs on rubric-based rewards spanning diagnostic and multi-turn clinical reasoning improves performance on HealthBench and MedXpert relative to standard reward schemes.
The 22-DoF hand targets in-hand manipulation for humanoids at an accessible price point.
Reported internal projections show cumulative negative free cash flow of $278B from 2026-2030, with $856B earmarked for compute tab-outpaces, sustaining enormous datacenter hardware demand.
Analysts question whether humanoids can do useful factory, warehouse and service work reliably and at an economically sensible cost, without suggesting China is stepping back from the sector.
A Jev-like family of small decision models from Jared Palmer, drawing ~370 upvotes on Hacker News.
Xiaomi's MiMo v2.6 model release hit the Hacker News front page with roughly 229 points.
The tool removes guardrails from language models, drawing roughly 217 points on Hacker News.
Cloudflare's Python Workers reached GA on its edge platform, ~155 points on Hacker News.
A Show HN tool that detects conflicting intent when multiple coding agents work the same repo in parallel.
Free means $0 per token today, not unlimited or unconditional: both platforms rate-limit free use and require a free API key, some OpenRouter free endpoints may train on your prompts unless you opt out in account settings, and the roster changes daily. Verified against each provider's own pricing data at generation time — but pricing changes without notice, so confirm it's still free on the provider's own page before you rely on it.
Source:
