Today’s lead · TechCrunch
Anthropic cuts live internet from internal agent evals after real-world misfires
Anthropic says internal eval agents exploited real sites, touched government systems, and even submitted a false police tip, so it has cut live internet access for evaluations. The practical lesson is blunt: agent evals need containment and monitoring before they touch public services.
Top signals
6 more
Cloudflare adds Clef-omni for multimodal decision models
Cloudflare is extending its Clef decision-model line with Clef-omni, adding native audio, video, image, and text inputs in one pipeline. The useful bit for builders is fewer brittle pre/post-processing chains; the claim to verify is whether decision quality holds across real multimodal payloads.
vLLM brings up NVIDIA Vera Rubin, with very early throughput and economics claims
vLLM says it now supports NVIDIA Vera Rubin, with early MiniMax M3 on AgentX results showing more than 7.8x GB200 throughput. SemiAnalysis adds broader economics claims, but builders should treat these as early hardware/software bring-up numbers.
SGLang details Rubin kernel work for Kimi K3 and Miles RL
SGLang’s Rubin post is more useful than headline benchmark shouting: it names attention, MoE, communication, and verification kernel changes for Kimi K3. The improvements are incremental end-to-end, but the engineering details show what early hardware enablement actually requires.
TRL v1.15 makes fused LM-head scoring the default for memory-efficient post-training
TRL v1.15 removes a common post-training memory cliff by avoiding full logits materialization for SFT, DPO, KTO, GRPO, RLOO, and distillation. For builders fine-tuning small models with large vocabularies, this is the kind of default optimization that matters more than a new API.
BudgetPix adapts image-generation compute to visual complexity at inference time
BudgetPix attacks a real image-model cost problem: uniform patch tokenization wastes compute on simple regions. The proposed adaptive quadtree tokenization lets one pixel-space diffusion checkpoint run at different budgets, with reported fidelity close to full-compute baselines at 25% compute.
Nace open-sources Drex 1.5, a 128K-context decision model
Drex 1.5 is being pitched as an open, under-10B-parameter decision model with a 128K context window and TypeSafe SDK compatibility. If the benchmarks hold for your domain, the interesting part is not leaderboard position; it is local, auditable decision scoring.
Tools & repos
4 selected
Busabase
Busabase pitches a shared system of record for humans and agents: records, docs, skills, apps, access control, and reviewable changes in one workspace. Interesting if your agents need memory without scattering state across tools.
Together Link
Together Link connects coding agents to open models on Together AI without forcing a workflow rewrite. The pragmatic value is cost control inside existing harnesses and tools.
alibaba/open-code-review
Alibaba’s open-code-review combines deterministic code-review pipelines with an LLM agent, including line-level comments and built-in rules for NPEs, thread safety, XSS, and SQL injection.
Zernio
Zernio wraps marketing channel integrations behind one API, MCP server, and CLI. Useful for agent-built products that need distribution plumbing, but the hard part remains trusting the platform approvals and maintenance claims.
Blogs worth your time
3 reads
Impactful scheduling for GPU clusters
Ai2’s scheduler write-up is a rare production view into GPU allocation politics: budgets, hierarchical fair-share, and time-slicing replaced priority games. The notable outcome is less squatting, lower debug latency, and 74% fewer human-in-the-loop repairs.
How to Build a Model Router in the Harness
LangChain’s model-router post is practical because it ties routing to traces, task mix, and outcome tracking rather than gateway magic. Their Open SWE test cut median thread cost 64% with no measured quality drop.
A new feature for my blog, built using my voice
Simon Willison’s voice-Codex experiment is a grounded look at where coding agents help today: rough feature construction while multitasking, then typed review for exactness. The workflow shipped, but it still needed human code review and corrections.
Community discussions
2 threadsUzu claims 92 tokens/s local Qwen3.5-9B on a 16GB Mac
The interesting tension is not “llama.cpp bad”; the post says ordinary decoding is memory-bandwidth bound. Uzu’s claimed speedup comes from Apple Silicon-tuned speculative decoding, so reproducibility and output quality are the checks.
H100 rental prices are rising instead of depreciating
The discussion is a compute-market warning: if H100 prices rise three years after launch, depreciation assumptions for GPU debt get shaky. Installed capacity gains leverage while demand keeps absorbing newer supply.
Funding & acquisitions
3 moves
TypeSafe AI raises $870M at a $7.5B valuation after Jev launch
TypeSafe’s giant Series A is a bet that decision models become automation infrastructure, not chat UX. The adoption claims are eye-catching, but builders should separate calibrated-decision utility from launch-week valuation heat.
Sabi raises $50M for a non-invasive brain-AI interface
Sabi’s $50M raise targets a wearable BCI cap for thought-to-prompt interaction. The ambition is clear; the supplied evidence gives investor names and product direction, not proof of capability.
CurvetAI raises Rs 6 crore pre-seed for multi-model AI agent workflows
CurvetAI is raising around the integration pain of generative AI: one workspace for text, image, audio, and video models, then agents to coordinate them. Early traction is claimed, but revenue details were not disclosed.
Bengaluru radar
2 events
Yapmaxxing: AI for productivity at work
Hands-on AI productivity masterclass in Bengaluru today: bring a laptop, build voice-driven workflows, and compete in the automation challenge.
QuadX: The Next Decade of AI with Kunal Dwivedi
Morning HSR session for students and builders: Google engineer Kunal Dwivedi discusses AI’s current state, career skills, and the next decade.