|
|
Hi there, this is your daily ☕️ DevShot.
|
|
|
In today's DevShot:
|
|
🌍 OpenAI launches GPT-6.1 Terra 🔀 Entire launches a GitHub alternative 🧩 JetBrains unveils self-customizing IDE 🐍 Microsoft unveils type server protocol 🔧 Ubicloud fixes vLLM tool calls Plus: 🎁 7 more stories you might like, 🛠️ 5 tutorials, 🧰 7 tools, and 📚 5 papers.
|
|
The AI newsletter curated by engineers who build itTLDR AI is a free daily email curated by Anthropic and ex-Google engineers: the new tools worth trying and the research worth understanding, summarized for 1.1M+ readers.
|
🌍 OpenAI launches GPT-6.1 Terra
LINK
|
- OpenAI yesterday shipped GPT-6.1 Terra, a faster, cheaper mid-tier model built for routine office tasks, replacing GPT-5.6 Terra and slotting below Sol and Astra but above Luna in the lineup.
- Its headline feature is computer use: Terra reads a screen and acts on it by clicking, typing, and switching apps, while cutting the error rate on a hard factual-question set from about 54% to 17% at the lowest setting.
- Developers get Terra today in ChatGPT paid tiers and Codex at $1 per million input and $5 per million output tokens, half the old price, though OpenAI still points to Sol or Astra for the toughest coding work.
|
🔀 Entire launches a GitHub alternative
LINK
|
- Entire, the startup from former GitHub CEO Thomas Dohmke, launched its full GitHub alternative yesterday, replacing the pull request with an agentic code review system called Trails built around a distributed Git-compatible server.
- Trails runs customizable agent-scored evals on every push, gauging risk or intent drift across long LLM sessions, surfacing findings in the web UI and an open-source CLI, and folding the original request plus agent session history into the review.
- The platform is free and you pay only for tokens the review agents consume via Claude or OpenAI models, with a $100 starter credit; private repos can be hosted in the US, Europe, India, or Australia, while public repos and self-hosting remain on the roadmap.
|
🧩 JetBrains unveils self-customizing IDE
LINK
|
- JetBrains has unveiled ThinkRail, an agentic IDE that learns your codebase through persistent "living specs" and reusable skills, with planned extension support that lets the agent build its own tools around how you actually work.
- It wraps a coding agent in a real workbench of Monaco editor tabs, terminals, and a git-worktree model where each isolated workspace gets its own branch, runs agents in parallel, and carries GitHub PR state with CI checks and review decisions inline.
- ThinkRail ships as a native desktop app or self-contained CLI for macOS 14+, Windows x64, and Linux, uses JetBrains AI or your own provider credentials with no accounts of its own, and never sends prompts, code, or files in its usage analytics.
|
🐍 Microsoft unveils type server protocol
LINK
|
- Microsoft has published the Type Server Protocol (TSP), a JSON-RPC protocol that lets clients query a Python type server for analysis data, including protocol version, snapshots, search paths, import resolution, and computed, declared, and expected types.
- TSP borrows LSP data shapes and runs over stdio or Node IPC, with an optional multi-connection mode that opens extra read-only channels for concurrent type queries while the main connection keeps ownership of workspace state and document synchronization.
- Returned type handles and IDs are bound to a snapshot that only advances when analysis results change, so requests carrying a stale snapshot should fail with LSP's ContentModified error, letting the client retry against a fresh snapshot.
|
🔧 Ubicloud fixes vLLM tool calls
LINK
|
- Ubicloud eliminated malformed tool calls when serving Kimi K2.6 and DeepSeek V4 Flash by enabling vLLM's constrained decoding, lifting schema accuracy on Moonshot AI's K2VV ToolCall benchmark from 92% to a full 100%.
- Constrained decoding, shipped as a server-side flag in vLLM 0.31 and backed by the XGrammar library, guides generation token by token, masking out any token that would break the tool's schema while adding negligible GPU overhead.
- vLLM keeps the feature off by default because per-model bugs hit models like gpt-oss-20b and Qwen 3.6, so engineers wanting first-party quality must opt in via the new `--tool-strict-level` flag and verify against the benchmark.
|
|
Watch Us Build Your First Agent Workflow LiveIf logging, enrichment, and routing are eating your team’s week, this one’s for you. We’ll build the agents that handle that go-to-market work in HubSpot and Dust, so your team can get back to selling. You’ll watch us build a post-call automation and a Slack account ownership layer. Then you can set up the same thing for your team.
📚 Tutorials & Advice
|
> Billion-query-per-second lookup tables in C++: Precompute a collision-free lookup function at compile time for fixed key sets, cutting string-to-value lookups to about five processor cycles and beating the standard C++ map thirteenfold.
|
|
|
🔗 Other News & Articles
-
Nightly-labs/Openbot: A local-first desktop workspace for persistent AI teammates. Run Codex, Claude, and Grok with dedicated workspaces, task queues, file sharing, browser control, and agent-to-agent collaboration.
LINK
-
Oldest GUI application that still runs on Windows 11
LINK
-
Accelerating HTTP web servers with in-kernel eBPF deliver up to 4.5x throughput
LINK
-
Trace Claude managed agents and the Messages API in Sentry
LINK
-
Microsoft.UI.Reactor: a C#-first way to build WinUI apps
LINK
-
OpenTofu 1.13 lets module authors reduce unknown values at plan time
LINK
-
Ollama for Mac gains support for local text-to-speech models
LINK
|
|
📄 Papers & Resources
|
> A decision-model factory builds custom, company-controlled models for specific tasks on demand, with five beating their starting points, three topping a Japanese benchmark, and a serving trick doubling speed at unchanged accuracy.
LINK
|
|
> Tool-use training data is generated by letting AI explore ~1,000 real business APIs and work backward from results, producing verified tasks that let a small model match Claude Sonnet 4.6 on tool-calling tests.
LINK
|
|
> Software documentation checks catch auto-written code descriptions that are technically true but no longer match what the code actually does, lifting verified-accurate statements from 83% to 95% and surfacing 37% more trustworthy explanations.
LINK
|
|
> AI agent conflicts get a new test showing that when an agent's stored memories clash with its skill instructions, it often picks a winner based on word order rather than meaning, and quietly resolves conflicts without telling users.
LINK
|
|
> Automated app migration to HarmonyOS rewrites Android or iOS apps into Huawei's new language and tests them on real devices, nailing app structure up to 100% but still faltering on live behavior as low as ~4%.
LINK
|
|
|
|
🧰 Latest Tools
| ElevenLabs: turns any script into studio-quality speech in seconds, with 5,000+ voices across 70+ languages. Free to start. GENERATE FOR FREE |
|
Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during cooldowns.
LINK
|
|
Kilo Code for JetBrains: open-source AI coding agent for JetBrains IDEs that builds and iterates across projects with flexible model choices.
LINK
|
|
Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative data management.
LINK
|
|
iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate and fix issues
LINK
|
|
Inferock: a local proxy for OpenAI, Anthropic, Gemini, and OpenRouter that tracks token usage, failures, and retries to catch billing overcharges.
LINK
|
|
Enter Pro: an AI-native platform for building and scaling apps that unites development and growth, so you avoid stitching together separate tools and services.
LINK
|
|
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
|
