Summary
Today's feed was NVIDIA's day. The single strongest signal is a cluster of roughly a dozen posts around the GTC Taipei keynote (Jensen Huang in Taiwan), announcing the RTX Spark consumer superchip, the Vera "CPU for Agents," the 550B Nemotron 3 Ultra model, and Cosmos 3 for physical AI. Keynote slides captured in the feed give hard specs: RTX Spark is a 1-petaflop FP4 Blackwell GPU with a 20-core MediaTek-built Grace CPU and 128GB unified memory, while Vera is an 88-core agent-tuned CPU claiming 40% lower loaded latency than x86. The second-strongest thread is agent security: Tsinghua's Chain-of-Authorization (claiming a 98.5% attack success rate cut to 0% by baking authorization into the reasoning chain), Anthropic's Zero Trust for AI Agents framework, and NVIDIA's SkillSpector skill scanner all landed in curated reposts on the same day. A third cluster pushes the Nous Research Hermes agent framework, which NVIDIA confirmed runs on the new RTX Spark via the OpenShell runtime. Standout non-cluster items include a MIT/Stanford/NYU/Princeton paper arguing AI often makes people feel more efficient without actually being faster, the recurring "binding constraint thesis" that an agent's ceiling is set by harness quality not model intelligence, the Guild local-memory tool for multi-agent coding, and Kilo Code's note that Opus 4.8 now costs the same as 4.7. A large volume of French and Los Angeles political posts in the account feed carried no AI substance and is skipped.
Posts
NVIDIA GTC Taipei keynote: RTX Spark, Vera, Nemotron 3 Ultra, Cosmos 3 (cluster of ~12) (@nvidia, @Scobleizer, @ns123abc · NVIDIA Newsroom). Jensen Huang's GTC Taipei keynote dominated the feed. The headline is RTX Spark, pitched as the first Windows PCs purpose-built for personal AI agents. A captured keynote slide lists the specs: a Blackwell RTX GPU at 1 petaflop of FP4 (4-bit floating point) AI performance, a 20-core Grace CPU custom-built with MediaTek, 128GB of unified memory at 600 GB/s NVLink, and the full NVIDIA stack (CUDA, TensorRT, NVFP4, RTX ray tracing, DLSS). NVIDIA and Microsoft co-built a native Windows experience for agents with new security primitives and a runtime called OpenShell. A second slide shows NVIDIA Vera, explicitly titled "CPU for Agents": an NVIDIA-custom Olympus core design, 88 cores and 176 threads with spatial multithreading, a 164MB L3 cache, 250-450W, and a claimed 40% lower loaded latency than x86 plus 3.4 TB/s core-to-core bandwidth, all tuned for the latency-bound orchestration workload an agent actually runs rather than raw throughput. A third slide announces Nemotron 3 Ultra (550B), whose benchmark table puts it top on long context (95% on Ruler at 1 million tokens, where Kimi K2.6 and GLM 5.1 cannot run past 256K) and tied-best on agent productivity, but behind GLM on long-horizon planning (33% vs 40%) and behind Kimi on coding (54% vs 67%); Scobleizer separately relayed a Nemotron variant pitched as "5x faster, 30% cheaper." Cosmos 3, NVIDIA's open omni-model for physical AI, was shown generating a sound-on Formula 1 dashcam video from a text prompt. See the full hardware writeup: NVIDIA GTC Taipei summary and today's digest.
Chain-of-Authorization bakes permissions into the reasoning chain (@alex_prompter). A curated repost summarizing Tsinghua University work argues every enterprise AI agent shares one structural flaw: an LLM treats all accessible data as equally fair to share, with no concept of ownership, and the permission layer on top is "just a prompt" that gets bypassed. Chain-of-Authorization bakes authorization into the reasoning chain itself, so the model reasons about whether the requester is allowed to see each piece of data as part of its chain-of-thought. The reported result is an attack success rate cut from 98.5% to 0%. This is tweet-level evidence with no paper text attached, but it fits today's broader agent-security theme. Wiki: Chain-of-Authorization.
Anthropic's Zero Trust for AI Agents framework (@_vmlops · PDF). A repost of Anthropic's zero-trust playbook for AI agents, which the poster frames as architecture rather than theory. It names threat classes traditional access controls were not built for: prompt injection through external data sources, tool poisoning via MCP server metadata, memory-based privilege retention across sessions, and multi-agent pivot attacks. The framework splits defenses into three tiers (Foundation, Enterprise, Advanced). The linked PDF body could not be parsed by the farmer, so this is from the post text. It pairs directly with Chain-of-Authorization and the ClawTrojan paper covered in today's digest; see ClawTrojan summary.
NVIDIA SkillSpector: a security scanner for agent skills (@bibryam · GitHub). A new NVIDIA open-source tool that scans AI agent skills before installation, running 64 security checks across 16 categories including prompt-injection detection, credential-theft detection, supply-chain vulnerability scanning, and AST plus taint-flow analysis, with an optional LLM review layer and SARIF output for CI/CD. The poster frames it as "Semgrep plus antivirus" for agent skills. This is the tooling complement to today's agent-security research cluster: the same week NVIDIA ships agents to consumer machines, it also ships a scanner for the skills they run.
Nous Research Hermes runs on RTX Spark (@ns123abc). Nous Research announced that Hermes Agent now runs on NVIDIA's RTX Spark superchip and integrates with the new OpenShell runtime, which connects Hermes to Microsoft's security primitives, featured in the Computex announcement. This is the substance behind the broader Hermes promotion in the feed: the open agent framework is now part of NVIDIA's consumer agent stack, which is why Scobleizer noted the keynote video literally said "running the Hermes harness."
Hermes Agent Control Room and masterclass promotion (cluster of 3) (@AlphaSignalAI, @cyrilXBT). A set of reposts promoting the Hermes agent framework. The Control Room is a public template that treats an agent fleet like an operating system, growing in four stages from one agent on a VPS to an orchestrator routing across specialists with cron jobs running the whole team. The cyrilXBT posts are a "zero to production agent in one sitting" masterclass. The underlying x.com article bodies could not be fetched (expired cookies), so this is from the visible post text; the substance is promotional rather than a research result.
MIT/Stanford/NYU/Princeton: AI's productivity feeling can outrun the reality (@rohanpaul_ai). A repost summarizing a new paper finding that AI can make people feel more efficient even when the measured benefit is tiny, missing, or negative. People reach for AI on simple tasks because it feels like it saves time, and the bigger danger is a feedback loop: once people use AI they become more likely to use it again, even for easy tasks where doing it themselves would be just as fast. The claim is that AI dependence can grow from a mistaken feeling of convenience rather than real productivity gains.
The binding constraint thesis: harness beats model (@IntuitMachine). A repost arguing the "better model equals better agent" assumption is wrong: frontier teams see GPT-5 still failing 60% of long coding tasks, but the same model with a better harness delivering a 10x improvement with no new weights. The framing is that an agent's ceiling equals the minimum of model capability and harness quality, and right now the harness is the binding constraint, a Ferrari engine in a go-kart frame. This is the conceptual opposite of the "compile the workflow into weights" paper in today's digest: one camp says keep improving the external harness, the other says dissolve it into the model.
Guild: shared local memory for multi-agent coding (cluster of 2) (@AlphaSignalAI · GitHub). A single Go binary backed by embedded SQLite that runs a local MCP server so Claude Code, Cursor, and Codex can share the same memory and hand off work cleanly, with nothing leaving the machine. Search blends BM25 keyword matching with vector similarity and fuses both. It addresses the practical problem that multi-agent coding setups keep losing context across tools and sessions.
Claude removed temperature controls, and so did everyone else (@alex_prompter). A repost pushing back on the community reaction to Claude removing manual temperature (sampling randomness) controls, noting OpenAI already locked temperature on o1, o3, and GPT-5, and Google warned against changing it on Gemini 3. The argument is architectural: reasoning models run internal chain-of-thought verification across multiple passes, and forcing temperature to zero would collapse those paths into one greedy line and sabotage the model's reasoning. Opinion rather than result, but a useful read on why frontier reasoning models are converging on fixed sampling.
Kilo Code: Opus 4.8 at 4.7 pricing, Step 3.7 Flash free (@kilocode). The coding-agent tool noted three arrivals sharpening the price-to-performance trade: Opus 4.8 at the same price as 4.7, StepFun's Step 3.7 Flash currently free in Kilo, and Xiaomi cutting MiMo pricing by up to 99%. A small but concrete data point on model pricing pressure.
AI Layoff Trap economics paper (@jackcoder0). A repost dramatizing a March 2026 Wharton and Boston University paper that models a scenario where firms automate toward boundless productivity and zero demand, an economy that produces everything and sells to nobody. The post is rhetorical, but the underlying paper is a peer-reviewed macro model of automation-driven demand collapse, worth noting as the economic-anxiety counterweight to the day's hardware optimism.
Opaque x.com article reposts (@cyrilXBT, @zodchiii, @ashwingop, @dair_ai). Four curated reposts point to X native long-form articles whose bodies could not be fetched because the session cookies were expired. The dair_ai one is the DAIR.AI Top Papers of the Week roundup, whose content was recovered separately from Gmail and fed today's digest (SkillOpt, Compiling Agentic Workflows into Weights, AutoScientists, Life-Harness). Click through to read the others.
GPT-Realtime 2.0 voice computer control demo (@Scobleizer). Scobleizer reshared a demo of controlling a computer entirely by voice via GPT-Realtime 2.0, calling it underrated and the future of operating systems. A demo video, no benchmark, but a notable data point on voice-driven agent interfaces arriving alongside the NVIDIA agent-PC push.
Political and crime account-feed posts (Skip). A large block from @brivael (French election commentary), @spencerpratt (Los Angeles politics), @AustinJustice, and @WHFraudTF carried no AI research or industry substance and is skipped.