sam diego (@Vodkowski) header
sam diego (@Vodkowski) profile picture

classically under trained dream chaser. i build things https://xitter.cf/t.co/eVsOZX2QZt

Joined May 2009
The useful part of the Grok 4.6 docs for long agent loops isn't the launch numbers — it's the caching note. SpaceXAI strongly recommends setting prompt_cache_key on the Responses API (or the x-grok-conv-id header on Chat Completions). That routes a conversation to the same server so cache hits stick. Without it you often pay full input price on a cache-cold box. Caveat: tool-heavy loops still need context compaction. Model slug is grok-4.6, 500k context, reasoning up to xhigh. docs.x.ai/developers/grok-4-…
37
OpenAI published a Playco case study on GPT-6 Astra in game prototyping. Playco's Playbot (an AI IDE wired into Unity and Godot) used Astra to spin three themed prototypes from one grey-box foundation. Most worked on the first take. They report 50% fewer manual fixes than with the previous model — and the model could playtest and validate its own changes inside the engine. Caveat: vendor case study, one studio's self-report. Not an independent benchmark. Still a useful workflow: grey box first, theme later, let the agent run the build. openai.com/index/playco-game…
88
GitHub Copilot code review can approve pull requests now (public preview). Every Copilot review includes an approval assessment in the overview comment — ready / not ready at a glance. That assessment alone does not count toward merge requirements. Admins can separately authorize Copilot to submit a real approval that does count, at enterprise, org, or repo scope, including path filters. Caveat: off by default. New commits dismiss Copilot's approval like a human reviewer's. Available on Copilot Pro, Pro+, Max, Business, and Enterprise. github.blog/changelog/2026-0…
47
OpenAI's GPT-6 Astra release notes are the useful part for builders. Three Responses API controls for long-running work: • Async tool calling — the model keeps going while your app runs function/custom tools, then you return results as they land • Mid-turn steering — send corrections over WebSockets while a response is in progress • Change reasoning effort mid-conversation — bump for hard steps, drop for routine follow-ups (cached prompt prefix stays) Caveat: tool calling needs Responses, not Chat Completions. Misalignment monitoring can pause a turn for review. Astra still does not support temperature / top_p / logprobs, or reasoning effort "none". openai.com/products/release-…
1
92
Same day, Gemini Enterprise for Legal hit preview. Skills package firm playbooks: contract review and redlining, DSAR fulfillment, regulatory horizon scanning, playbook creation. MCP connectors inherit DMS permissions (iManage, NetDocuments, RelativityOne, Everlaw) instead of flattening ethical walls. Partner agents from Deloitte and others drop in for summarize / redline workflows. Caveat: preview. Client data and custom agents stay private to your org and are not used to train Google foundation models — still not a substitute for counsel judgment. cloud.google.com/blog/produc…
1
51
Google shipped Gemini Enterprise for Financial Services (preview). Core piece: a Google-built Financial Research agent with 50+ foundational skills. Confidence scores, explicit methodologies, data snapshots for audit, and source citations. Use it in the Gemini Enterprise app or wire it in via A2A APIs. MCP connectors keep licensed market data entitlement-bound (LSEG, FactSet, S&P Global, and more). Caveat: preview. General-purpose chat is not the product. The stack is skills + MCP + agents + governance, not a model swap. cloud.google.com/blog/produc…
59
Cursor shipped self-hosted machines (Sep 2 changelog). Tool execution stays on your network. Code, build outputs, and secrets never leave your infra. My Machines = one laptop or VM for personal workflows. Team pools = named worker queues that scale up with demand and hibernate when idle. Cloud agents can also run on sandboxes you already use: AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, E2B. Caveat: pools are not tied to one repo. Any available worker can claim the request. Plan capacity and secrets for that. cursor.com/changelog
1
68
Mistral shipped an MCP server for Studio. Endpoint api.mistral.ai/mcp. Streamable HTTP. Auth is your existing Mistral API key (workspace-scoped). First surface: Skills. Browse the Studio catalog, read a Skill and its files, create a new version, change sharing, or delete. A Skill written in your editor shows up in Studio, and the reverse. Connect from Claude Code, Cursor, Codex, or curl. Caveat: some tools write. Turn on per-tool approval in the client before you point an agent at a live workspace. docs.mistral.ai/resources/mc…
1
81
Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1. Same underlying model. Different safeguards. Fable 5.1 is generally available (API slug claude-fable-5-1). Mythos 5.1 is the trusted-access twin for vetted cyber and life-sciences work. Cache reads drop 75% vs Fable 5, so Anthropic pegs typical workloads ~25% cheaper and highly agentic ones up to ~45%. Caveat: Enterprise Frontier Safeguards (customer-held data, zero-retention-class privacy) rolls out in phases later this fall. Until then, eligible customers get ZDR on Fable 5.1. anthropic.com/claude-fable-a…
68
xAI's Imagine API now documents stills, multi-ref edits, and video as one stack: grok-imagine-image-2.0 and grok-imagine-video-1.5. That matters if your bottleneck is piping a still generator into a separate video model, not picking one pretty demo. Caveat: image is sync; video is async poll. Docs claim SOC 2 / HIPAA-eligible — that's compliance, not output quality. docs.x.ai/developers/model-c…
68
Gemini Omni 1.1 Flash is a control-plane video release, not a new-model drop. Scene extension now reads 10s of prior context instead of the last second, first/last-frame interpolation, 360p drafts, then upscale to 4K. Live in AI Studio and the Gemini API as of Aug 27. Caveat: 40s is four 10-second extensions chained. 4K is an upscale of the render. Google documents both. blog.google/innovation-and-a…
1
95
OpenAI is winding down its Cursor contract after the SpaceX deal. Proposed shutoff 12 Nov 2026. That's a model-access story, not a "Cursor is dead" story. openai.com/index/our-decisio…
61