SpaceX Starship Reaches Orbit as Anthropic, Perplexity and Hugging Face Push New Agent Infrastructure

Key Takeaways

1. SpaceX's Starship completed its first orbital flight, deploying Starlink V3 satellites and marking the third-largest payload mass ever launched to orbit.
2. Anthropic released Claude Sonnet 5.5 with over 30% faster speed and up to 30% lower cost for most workloads.
3. Perplexity added Profiles, Skills and managed connectors to its Agent API, while Hugging Face redesigned Gradio and shipped new agent tooling.
4. NVIDIA and Andrew Ng advanced open agent safety infrastructure, with OpenWorker built on Nvidia OpenShell to sandbox agent commands.
5. Runway integrated ElevenLabs v4 for narration and character dialogue, and Grok Bot launched team features plus Grok 4.7 on Amazon Bedrock.
6. OpenAI and Anthropic published guidance on frontier RL training safety, while debate continued over alignment research and AI's labor-market impact.

1. AI Models and Infrastructure

1. Anthropic released Claude Sonnet 5.5, the second model in the Claude 5.5 series, claiming a clear upgrade over Sonnet 5 with over 30% faster operation and up to 30% lower cost in most work scenarios. Early observations from Factory note that High is a strong default tier, the model checks whether real requirements are met rather than just passing the nearest failing test, and it questions explanations carried over from earlier work. 1 2
2. Runway now offers ElevenLabs v4, a model the company says excels at narration and character dialogue with expressiveness and tonal naturalness beyond previous versions, usable alongside Runway's image and video models. 1
3. Grok Bot introduced Team Bots with configurable skills, plugins and credentials, one-click Slack joining and multi-person collaboration; it can also assist with finance, business dashboards and subscription savings. Grok 4.7 is now available on Amazon Bedrock, and a cited post ranks Grok 4.7 xHigh first on the Artificial Analysis Cyber Index. 1 2 3 4 5

2. Agent Platforms and Developer Tools

1. Perplexity Agent API added Profiles, Skills and managed connectors. Profiles are saved agent configurations (model, instructions, tools, connectors, run settings) configured once by project admins under a versioned ID and callable by project developers with a project API key; Skills are versioned so flows like v2 can ship without changing integrations; connectors initially support GitHub, Slack, Drive, Datadog, Linear and Notion, configured by admins so apps do not hold credentials. Profiles and Skills are live, connectors are in preview. 1 2
2. Hugging Face is redesigning Gradio from scratch after five years, saying building frontends for AI apps is no longer the main obstacle, and is collecting feature requests. It also highlighted Trackio 0.39, which lets embedded dashboard agent apps read the current dashboard view and query only that portion, and Holo4, a new batch of general computer-use models that can click, write code and call tools across desktop, web, Android, code sandboxes, MCP servers and enterprise APIs, commercially available on H Models API. 1 2 3
3. Hugging Face Jobs usage grew via agents: with only the hf CLI installed, one test of hundreds of code snippets on a Hub model page launched 322 jobs in 90 minutes, up to 25 concurrent, spanning CPU to A100 and diffusers, transformers, sklearn, MLX and Keras, for about $4 total. 1

3. Agent Safety and Evaluation

1. NVIDIA and more than 100 industry partners launched the NVIDIA Open Agent Safety Platform, combining OpenShell and Sentry as an open-ecosystem starting point for a trust layer for safe agent systems. Andrew Ng said the OpenAI-Hugging Face attack stemmed from weak sandboxing and called Nvidia's open-source agent sandbox tooling a positive step; his OpenWorker, built on Nvidia OpenShell, will run each agent's commands in a sandbox with only task-relevant files, no default access to keys, browser login credentials or arbitrary websites, limits enforced by deterministic code rather than error-prone, prompt-injection-susceptible LLM prompts, and all operations logged for monitoring and audit. 1
2. OpenAI published its thinking on safety protections for frontier reinforcement learning training runs, and Greg Brockman amplified a practical guide on the same topic as reflecting current lessons learned. 1 2
3. Hamel Husain reviewed Anthropic's new evaluation tool: drawbacks include writing evals before looking at data, using markdown files instead of a purpose-built annotation app, overly broad eval scope bundling multiple failure types, and getting stuck in individual eval details too early; strengths include the strongest out-of-the-box issue discovery among comparable auto-evals, catching handoff, formatting and voice-agent issues, much improved UX over prior eval plugins, and blog views on looking at data, smart sampling and avoiding eval saturation that he agrees with. He also noted eval experience is becoming a common PM job requirement, citing Ramp's receipt capture accuracy rising from 35% to 83%, Shopify's AI workflow builder being 2.2x faster and 68% cheaper than the frontier-model approach it replaced, Harvey nearly doubling internal quality scores after rebuilding its AI contract reviewer, and Cursor's Auto Balance routing cutting costs 41% while improving satisfaction. 1 2

4. Industry Debate and Research Signals

1. Yann LeCun amplified a long critique arguing 15 years and billions of dollars of alignment research have produced no substantive progress, claiming its leaders do not understand computers or large-scale distributed systems, do not know the inference-serving stack, cannot uniformly define "agent," and presuppose an "everyone will die" conclusion that makes any model behavior read as bad news. He also amplified criticism of Anthropic's hired AI ethics philosopher as forming a closed subculture akin to "bioethics" practitioners who favor clever arguments over possible human suffering, and questioned whether Anthropic's AI truly made scientific discoveries independently, comparing it to a prior Navier-Stokes solution coincidence. 1 2 3
2. On labor-market impact, LeCun amplified an NBER study by Jane Wu and Robert Fairlie finding AI has not taken jobs from new college graduates. Ethan Mollick said AI's employment effect remains very unclear, with employers still figuring out individual-level adoption, let alone use in complex organizations, and change may accelerate as agents capable of real work spread; a cited post synthesizing about 20 papers found no aggregate labor-market shock, with unemployment and layoffs nearly unchanged, but possible cuts to AI-exposed white-collar entry-level hiring, with remote work possibly explaining part of the decline. 1 2
3. Ethan Mollick argued AI construction exceeds the railway-building peak as a share of GDP but dominates the economy far less: in 1890, 1 in 12 American men worked for railroads, which used 56% of all machine horsepower. He also said Apple's choice of an on-device, capability-limited direction for Siri may be wrong as cloud-based "Claw-type" assistants with powerful virtual computers become a personal AI direction, and noted last year's DevDay description of agent form was abandoned quickly after OpenClaw and self-organizing swarms emerged, showing AI labs cannot always predict how fast their own products change. 1 2 3

5. Funding, Acquisitions and Community

1. Workera was acquired by Pearson. Andrew Ng said Kian Katan judged in 2019 that rigorous skills measurement would become important, a judgment AI progress has made more valid; Workera's skills measurement helps companies understand employee strengths and development areas, and under Omar Abbosh and Kian the combined entity could serve more people. Kian said Workera blends AI, psychometrics and enterprise execution expertise, pioneered AI-native skills intelligence, agent-led multimodal assessment and environmental skills measurement, and built cross-organization, industry and role skill benchmarks; he said AI is changing work, with some roles disappearing and new ones appearing, requiring help for billions to develop new skills. 1
2. Yann LeCun amplified a call for researchers in continual learning, multi-agent parallel work, and synthetic data, environments and evaluations for measuring frontier capabilities, saying there is enough funding for new research and a willingness to share results publicly. He also amplified an October 10 San Francisco reading group on world models, sharing his ICML paper on world-model representation learning and recent work AdaJEPA, and a public letter to the UK government calling for an end to unfair restrictions blocking UK talent from changing jobs or starting companies, backed by UK AI companies that have raised $5 billion. 1 2 3
3. swyx amplified an interview in which Anthropic's Thariq Shihipar explained prompting remains one of the highest-leverage skills in agentic coding, Claude.md may eventually disappear, Claude Mods lets developers customize the entire Claude Code harness, mutable software may change how apps are built, multiplayer agents and Claude Tag are becoming organization-level harnesses, and agent attacks on Hugging Face expose a larger security problem that worsens as frontier models gain capability. 1