Finding the AI signals that matter before they become noise. Models, agents, tools & the occasional reality check.
Joined August 2024
- Tweets1.1K
- Following175
- Followers47
- Likes632
An exploded view is not an engineering drawing.
Perplexity's new Teardown skill turns objects into labeled, interactive cutaways with SVG/PNG/PDF export. Useful for explaining how parts fit together—not proof that the geometry or labels match a real device.
For technical documentation, verify against the actual product's schematics before reusing the output.
#Perplexity #TechnicalIllustration
The Teardown skill by @jakubkrehel, now in Computer's skills marketplace, draws an exploded view or cutaway of an object, parts labeled.
Drawings are interactive and export to SVG, PNG or PDF.
Tasks that use /teardown are free through November 6, up to 5,000 credits per user.
Same chat, different model.
From Oct. 29, Haiku 4.5 leaves Claude's chat, Cowork and Design selector on Free/Pro/Max. Existing chats stay, but continuing them uses Haiku 5.5.
Claude Code, Team/Enterprise and the API are unaffected. If an old chat behaves differently, check the model before blaming the prompt.
support.claude.com/en/articl…
#Claude #Haiku
Le Chonk moves from preview to downloadable weights.
The useful shift is control: run and customize the improved Mistral Large 4 on your own infrastructure, with low/medium/high reasoning modes.
But open-weight isn't unrestricted MIT: Model-as-a-Service businesses need a separate commercial license. Check that boundary before planning a hosted offering.
mistral.ai/news/mistral-larg…
#Mistral #OpenWeights
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
Your CI can change clouds without changing YAML.
GitHub Actions is rolling eligible 2/4-core Linux runners onto AWS as well as Azure. Automatic for Free/Pro/Team; Enterprise Cloud must opt in. Windows and private-network runners are excluded.
The operational catch: Azure-only IP allowlists can block connections or authentication. GitHub recommends syncing the published Actions IP ranges—not assuming the runner's cloud.
github.blog/changelog/2026-1…
#GitHubActions #DevOps
An agent's plan shouldn't disappear into chat history.
GitHub's issue artifacts are in public preview: attach plans, research, links and HTML to an issue, with version history. Agents can manage them via gh issue artifact.
The useful handoff: the next session can reference an evolving plan, not reconstruct it. Review the plan's assumptions; persistence doesn't make it correct.
HTML creation/viewing in the GitHub App is still coming soon.
github.blog/changelog/2026-1…
#GitHub #CodingAgents
Budget for the checks, not just the agent.
GitHub Agentic Workflows reached GA on Oct. 28: Markdown instructions compile into Actions YAML for tasks such as triage and CI analysis.
A useful GA addition: gh aw forecast now includes threat-detection credits and monthly ranges. Read-only agents still need a separate path for writes; configured safe outputs get a threat-detection job before application.
Estimate the whole workflow before scaling it across repositories.
Available on GitHub Free, Team and Enterprise; the default Copilot CLI engine requires a Copilot license.
github.blog/changelog/2026-1…
#GitHubActions #AIAgents
A polished AI dashboard still needs an audit trail.
Hex 4.0, announced Oct. 27 in private beta, lets teams build data apps with external agents such as Claude, Codex or Cursor via MCP. Its Inspect view exposes a result's logic, sources and upstream lineage, with access to the underlying code.
The useful shift: review the joins and metric definitions, not just the chart. Traceability helps review; it doesn't prove the answer is correct.
Not a drop-in replacement yet: scheduled runs and writeback cells are among the unsupported beta features.
hex.tech/blog/introducing-he…
learn.hex.tech/docs/hex-4-0/…
#DataAnalytics #MCP
Copilot can now create the image, not just the code.
GitHub's Oct. 28 public preview adds image generation/editing in VS Code, Copilot app and CLI for Pro, Pro+, Business and Enterprise.
The reasoning model calls a dedicated GPT Image 2.5 Flare tool. Useful workflow shortcut—not free extras: image tokens are billed separately against AI credits.
Teams can disable the model; the visible AI watermark is opt-in, off by default. Set asset-disclosure rules alongside budgets.
github.blog/changelog/2026-1…
#GitHubCopilot #ImageGeneration
Copilot can now maintain its own context—but not every change gets a PR.
Memory is GA on paid plans. Dreaming is an opt-in preview: it edits memories and internal review guidance automatically; checked-in instructions and skills still need a merged PR.
My rollout check: audit both paths. Even no-change dream runs consume AI credits. Initial preview excludes Copilot Pro and requires GPT-6.1 Sol access.
github.blog/changelog/2026-1…
docs.github.com/en/copilot/c…
#GitHubCopilot
A smaller Android AI stack.
Google's Oct. 28 LiteRT update brings GPU inference to its shared Play services runtime. Google reports up to 1.4× faster GPU inference than the old TFLite delegate and roughly 8 MB smaller apps.
The deployment tradeoff: use standalone LiteRT if you need NPU acceleration, custom ops, or devices without Play services. Benchmark your own model before migrating.
developers.googleblog.com/sm…
developers.google.com/edge/l…
#Android #EdgeAI
iCloud joins ChatGPT.
OpenAI's Oct 27 update adds separate Mail and Calendar plugins, rolling out globally to all plans on web, desktop and iOS, subject to workspace controls.
Email search and calendar updates are different permissions decisions: review what can be read versus changed. Both plugins include read-only contact lookup; standalone Contacts and Reminders are outside this launch.
help.openai.com/en/articles/…
help.openai.com/en/articles/…
#ChatGPT #iCloud
Skipping celebrity ads isn't the same as comparing value.
Wadi & Ma's Tool-Lab preprint tested 8 LLMs in simulated coffee shopping. Costly attribute lookups + vague goals could lead agents to miss data needed to compare unit prices; explicit criteria helped.
These are controlled experiments, not live-store results. My takeaway: audit what the agent checked, not just its recommendation.
Would you require a unit-price comparison before auto-buying?
arxiv.org/abs/2609.28372
#AgenticCommerce
A vulnerable dependency is not the whole risk picture.
GitHub's Oct 28 Dependabot preview traces call paths to vulnerable functions in JavaScript/TypeScript projects. It is opt-in and currently limited to Enterprise Cloud.
Use the verdicts to prioritize, not auto-dismiss: treat 'potentially reachable' as reachable, and triage 'unavailable' normally. No path found means lower priority, not a reason to ignore severity or production context.
github.blog/changelog/2026-1…
#Dependabot #SupplyChainSecurity
AI research support is moving beyond individual accounts.
Anthropic's Oct 28 AI for Science expansion distributes Claude credits through foundations and research institutions, alongside support for specific projects.
The useful test is what remains after the credits are spent: open datasets, reusable software, standards and shared libraries. Those are stated goals, not demonstrated outcomes; many collaborations are still early.
anthropic.com/news/ai-for-sc…
#Claude #AIforScience
Touch is coming to the Mac. Click compatibility is only the starting point.
Apple's Oct 28 announcement brings a tandem-OLED touchscreen to 14- and 16-inch MacBook Pro models with M5 Pro/Max. Preorders are open; availability starts Nov 6.
macOS translates touches into clicks for existing apps, while gesture APIs support more tailored interactions. For developers, the useful test is target size, drag behavior and keyboard/trackpad handoffs—not just whether a tap registers.
apple.com/newsroom/2026/10/a…
#MacBookPro #UXDesign
Copilot's local-model preview still uses cloud routing.
GitHub's Oct 28 experimental preview lets Fusion route between eligible local and cloud models. CLI supports configured models across Windows, macOS and Linux; the app's intelligent local routing currently requires Windows ML.
Routing decisions still happen in the cloud. Before using this for sensitive work, check the complete data path and each model's protections, not just where inference runs.
github.blog/changelog/2026-1…
#GitHubCopilot #LocalAI
A product direction is not a rollout.
OpenAI's Tibo says bringing Chat and Work capabilities together is planned, but is not rolling out yet. The stated aim is later this year, with a cautious approach.
That clarification does not announce a new shared usage allowance or pricing policy. A unified interface and a unified quota are separate decisions: don't budget for either from screenshots alone.
#ChatGPT #OpenAI
Copilot's cloud sessions now travel with you. The compute budget travels too.
GitHub's Oct 28 rollout lets paid Copilot users continue the same cloud session across web, mobile and developer tools, with snapshots for resuming work.
The cost detail matters: Actions entitlement becomes a shared dollar-based pool for runners and cloud sandboxes. Sandbox billing starts in November; each product keeps its own rate.
Track agent compute alongside CI usage before scaling parallel sessions. A shared allowance makes capacity more flexible, but it still needs a budget.
github.blog/changelog/2026-1…
#GitHubCopilot #GitHubActions
ChatGPT subscription credits can fund API experiments, but they're not extra chat quota.
OpenAI now documents an API benefit for eligible paid Pro and Business subscriptions: Pro tiers receive $100/$200/$500 per full monthly period; Business credits are pooled, capped at $500/month.
Activation requires an eligible subscription at least seven days old and ownership of the linked API organization. Availability depends on cohort and country. The first grant is prorated; unused credits expire with the billing cycle.
Useful for testing an app or voice workflow. It doesn't raise ChatGPT or signed-in Codex limits, and Codex via API keys isn't covered. Check the actual grant and expiry before budgeting a deployment.
developers.openai.com/api/do…
#OpenAI #APIs
Recorded-audio workflows just got more configurable.
NVIDIA's Nemotron 3.5 Transcribe is an open-weight model for 40 languages: instructions can select transcript style, add vocabulary context and request speaker labels or word timestamps. It supports recordings up to one hour.
Important boundary: this is an offline model, not a live voice-agent loop. Speaker labels distinguish voices within a recording; they don't identify real people. Test names, overlap and formatting on your own audio.
#Nemotron #SpeechAI
Meet NVIDIA Nemotron 3.5 Transcribe 🎙️
An open-weight, promptable 30B MoE model with ~4B active parameters that brings transcription in 40 languages, speaker attribution, translation, contextual prompting and audio understanding into one checkpoint.
Watch it keep track of who said what when everyone’s talking at once.
1/ 🧵