Today’s lead · Hugging Face Blog
Liquid AI ships a small on-device agent model with real tool-use ambitions
Liquid AI’s LFM2.5-2.6B is aimed squarely at local agents: tool calling, multi-step workflows, 128K context, and claimed strong instruction/tool-use scores in a 2.6B model. The practical hook is cost and privacy on commodity hardware; the caveat is that coding still trails larger models.
Top signals
5 more
NVIDIA opens Alpamayo 2 Super for AV reasoning, planning, and auto-labeling
Alpamayo 2 Super packages trajectory generation, Chain-of-Causation traces, VQA, 2D grounding, meta-actions, and auto-labeling into one 34B AV model. For autonomy teams, the value is workflow consolidation and inspectability, not just benchmark wins. Weights are on Hugging Face and inference notebooks are on GitHub.
FLUX 3 Video arrives, but the strongest claims still need daylight
Black Forest Labs’ FLUX 3 Video is showing up through Krea and OpenRouter, with claims of a unified architecture across video, audio, image, and action prediction. The release is reportedly gated early access, not open source today, so treat performance comparisons and open-weight promises as pending validation.
Anthropic’s compute appetite now reportedly includes a $10B Volta deal
TechCrunch reports, citing Bloomberg, that Anthropic signed a six-year, $10B cloud compute deal with Volta. The planned Norway data center, developed with Bitdeer, is expected to deliver 133 MW and use NVIDIA Vera Rubin systems. It is another signal that model roadmaps are compute roadmaps.
npm worm turns package install scripts into a credential-exposure event
The Keyv-linked npm incident is the kind of supply-chain compromise builders actually hit: preinstall scripts in dev and CI, token theft, and republishing machinery. The tricky response detail matters: SafeDep says remove the malware’s credential-revocation watcher before rotating exposed tokens, because revocation can trigger attacker code.
Google’s ADK repo shows why agent workflows need real trust boundaries
Google deleted three ADK Python repo workflows after Pillar Security showed a public GitHub issue could prompt-inject a triage bot into triggering a privileged fixer. The reported flaw was repository automation, not the ADK package itself, but the lesson is broad: bot identity is not authorization.
Tools & repos
3 selecteduber/ADR
Uber’s ADR is a Python repo for securing enterprise AI agents with observability, security benchmarking, and threat detection. The description says it is deployed at Uber.
obra/superpowers
superpowers describes itself as an agentic skills framework and software development methodology. It is Shell-based and unusually large on stars, so inspect before assuming conventional repo signals.
ZapDigits MCP
ZapDigits MCP connects Claude and ChatGPT to marketing data sources including Google Analytics, Search Console, Meta Ads, and 30+ sources. Useful if your agent workflows need live marketing context.
Blogs worth your time
3 reads
Simon Willison’s LLM 0.32 turns the CLI into a more agent-shaped tool
LLM 0.32 adds visible reasoning traces, OpenAI Responses support, server-side tools, structured streaming events, and content-addressable SQLite logs. The interesting bit is architectural: the CLI is becoming a practical substrate for tool-using agents.
LangChain’s voice-agent eval framework separates correctness from caller experience
LangChain argues voice agents need separate evals for execution, outcome, and experience. That framing is useful: a call can follow instructions and still fail the user, or resolve the task while feeling painfully unnatural.
NVIDIA makes the case for World Action Models over VLAs in robotics
NVIDIA’s post argues robot policies need world dynamics, not just semantic VLM backbones. Cosmos 3-based World Action Models promise better physical generalization, fewer task-specific demonstrations, and deployment tiers from workstation serving to Jetson Thor.
Community discussions
4 threadsClaude Code refusal behavior may change when the same request arrives as an image
A Claude Code user says a direct piracy-stack request was refused, but a screenshot of a similar setup led Claude to recommend and build it. The thread’s builder takeaway is inconsistency: multimodal context may route policy interpretation differently.
Enterprise RAG backlash: start with the questions, not the vector database
A builder who has sold RAG systems argues many enterprise “document AI” projects are really data ownership, structured-query, or cleanup problems. The useful heuristic: write the 20 target questions and answer five manually before approving the architecture.
A non-coder ships a Steam demo with Claude, but not by outsourcing judgment
The strongest part of this Claude Code story is not “AI made a game.” It is the workflow lesson: Claude built code, pipelines, and knobs, but the human still owned logic, taste, debugging, and final tuning.
Builders are debating whether frontier agents are optimized to overrule you
The post blames annoying frontier-model behavior on RLVR, refusals, filters, and long-horizon agent optimization. The thread is uneven, but the real tension is familiar: builders want capability without models deciding when to refactor, refuse, or improvise.
Funding & acquisitions
4 movesHappyRobot raises $150M Series C for enterprise operations agents
HappyRobot’s $150M Series C gives enterprise voice-and-workflow agents another large proof point. The company claims 150+ enterprise customers, 5x growth since Series B, and expansion beyond logistics into insurance, energy, telecom, and airlines.
Kily raises Rs 30 crore to automate digital-commerce operations for brands
Kily’s Rs 30 crore round backs a narrow but real enterprise-agent wedge: managing marketplace signals, decisions, and workflows for brands across e-commerce and quick commerce. It claims ITC among tied-up consumer brands.
Superleap raises Rs 36 crore for an AI-native CRM push
Superleap raised Rs 36 crore led by Peak XV’s Surge to build AI-native CRM for large revenue teams. The company claims $2M ARR, 10x annual growth, and enterprise customers including Razorpay, Aakash, Cars24, and MediBuddy.
NYAI raises $1.5M seed for India-focused legal AI infrastructure
NYAI’s $1.5M seed round is aimed at legal AI where verifiability matters more than generic generation. The Pune startup emphasizes citation-backed research, audit trails, on-prem deployment, and Indian statutory, regulatory, and judicial data.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.