Tag
106 notes collected under this theme.
Archive
Logged at IST: 2026-07-16 02:06 IST What it is: Mira Murati announcing Thinking Machines’ first model, Inkling, and pointing to the launch post. Gist: The important part is not just “open weights.” Inkling is a 975B total / 41B active multimodal Mixture-of-Exp…
Logged at IST: 2026-07-16 01:33 IST What it is: Boris Cherny arguing that agent-era engineering leverage still comes from automation, but now automation also includes encoded domain knowledge like CLAUDE.md, review rules, skills, and docs Gist: The core claim …
Logged at IST: 2026-07-15 22:23 IST What it is: Arvind Narayanan pointing to his ICML 2026 annotated keynote slides and highlighting new pushback on recursive self-improvement assumptions Gist: Narayanan’s frame is that the "AI as normal technology" view still…
Logged at IST: 2026-07-15 16:58 IST What it is: Bilawal Sidhu video titled “Fable 5 Is Insane. I Vibe Coded Terminator Vision.” Gist: From the visible YouTube description and page metadata, this is a build/demo video about creating a browser-based range-analys…
Logged at IST: 2026-07-15 14:02 IST What it is: Zhengyao Jiang claiming the first experimental evidence of recursive self-improvement in an autoresearch agent Gist: The specific claim is not generic "agents got better with more tuning," but that an agent spent…
Logged at IST: 2026-07-15 13:30 IST What it is: Ayush Paul’s writeup on prompt-injecting Claude’s memory and browsing system into exfiltrating personal data Gist: The attack chain was not about breaking the memory store directly, but about combining long-lived…
Logged at IST: 2026-07-14 20:54 IST What it is: Martin Fowler sharing Unmesh Joshi’s article on DSLs and LLM reliability Gist: The article’s core claim is that LLMs become much more reliable when they are constrained by domain abstractions and DSLs instead of …
Logged at IST: 2026-07-13 19:31 IST What it is: X post by antirez linking his blog post "Control the ideas, not the code" Gist: antirez extends the earlier X-thread argument into a full workflow claim: if you own the ideas, design, testing, and QA of a system,…
Logged at IST: 2026-07-13 16:04 IST What it is: X post from the geohot archive linking George Hotz’s blog post "I love LLMs, I hate hype" Gist: Hotz argues for a strongly pro-AI but anti-hype position: LLMs, coding agents, and related tools are genuinely usefu…
Logged at IST: 2026-07-13 12:19 IST What it is: X post by Mario Zechner adding nuance to antirez’s AI-code ownership point Gist: Zechner’s point is narrower and more practical than a general plea for control: if you control the types and interfaces, the rest o…
Logged at IST: 2026-07-13 07:59 IST What it is: Satya Nadella’s X article "The Reverse Information Paradox" Gist: Nadella argues that AI flips Arrow’s classic information paradox: enterprises now pay not only with money for intelligence, but also with propriet…
Logged at IST: 2026-07-13 07:49 IST What it is: X post by antirez on the "don't look at the code" debate in AI-coded systems Gist: He distinguishes between two very different ways an AI-coded codebase can come into existence: one where the human still controls…
Logged at IST: 2026-07-13 07:37 IST What it is: X post by Mario Zechner recommending Terry Tao’s post "Old and new apps, via modern coding agents" Gist: Tao describes using modern coding agents to port his old Java applets to JavaScript and revive them quickly…
Logged at IST: 2026-07-12 13:18 IST What it is: X post by dogfiles linking the blog post "Vim of Coding Agents" Gist: Frames Pi as the Neovim of coding agents: a minimal, hackable foundation that adapts to your workflow instead of forcing you into an opinionat…
Logged at IST: 2026-07-11 01:15 IST What it is: X post by Shantanu Goel summarizing Deep SWE 1.1 model-cost comparisons Gist: Claims GPT 5.6 Sol medium outperforms Opus 4.8 max at roughly one-sixth the cost, while GPT 5.6 Sol High performs similarly to Fable 5…
Logged at IST: 2026-07-11 02:17 IST What it is: X post by Rohan Verma linking his blog post "GPT-5.4 with Pi 0.69.0 is just nice" Gist: Argues that an agent harness stack getting boring is a success condition, not a failure. The post frames Pi 0.69.0 + GPT-5.4…
Logged at IST: 2026-07-11 15:38 IST What it is: Lilian Weng blog post, "Harness Engineering for Self-Improvement" Gist: Argues that recursive self-improvement in the near term is less about models rewriting their own weights and more about improving the surrou…
Logged at IST: 2026-07-11 01:09 IST What it is: X post by Mitchell Hashimoto on side-by-side Sol xhigh versus Ultra runs Gist: Says two days of side-by-side planning and implementation runs did not reveal a tangible quality difference between Sol xhigh and Ult…
Logged at IST: 2026-07-10 12:11 IST What it is: X post by Geoffrey Huntley linking to Stack72's essay "The Great Divergence in Software Engineering" Gist: Argues that the gap between teams effectively using AI and teams still piloting or rejecting it is no lon…
Logged at IST: 2026-07-10 00:05 IST What it is: X post by h100envy summarizing a long talk by the ex-NVIDIA engineer behind Unsloth on fine-tuning and reasoning-model workflows Gist: Frames a practical single-GPU stack for local/post-training work: choose a ba…
Logged at IST: 2026-07-10 00:04 IST What it is: X post by Steren announcing the public launch of Cloud Run sandboxes Gist: Claims Cloud Run sandboxes can start, execute, and stop 1,000 sandboxes in 5 seconds with roughly 500 ms average latency, positioning the…
Logged at IST: 2026-07-08 23:02 IST What it is: X post by Maksym Andriushchenko linking to a Substack post, "How I Use Codex To Automate Parts Of My Research Workflow" Gist: A pragmatic writeup on using Codex to reduce friction in AI safety research by offload…
Logged at IST: 2026-07-08 22:40 IST What it is: X post by Bilgin Ibryam linking to Prime Radiant's "Some new agentic patterns" Gist: Describes production-ish internal agent patterns built around an "agentic user in the loop" model, with agents in Slack handlin…
Logged at IST: 2026-07-08 22:32 IST What it is: X post by Steve Francia linking to his post "Why TypeScript 7.0 Was Rewritten in Go (and what it means for your dev stack)" Gist: Argues the TypeScript team’s Go rewrite is a broader signal that agentic software …
What it is: Drew Breunig revisiting his "10 Lessons for Agentic Coding" list and asking for additions Gist: the piece frames coding agents as making code cheap but not making judgment cheap; strongest lessons are to implement/rebuild to learn, invest in end-to…
What it is: hands-on writeup of deploying Cloudflare’s official agentic-inbox to run a custom-domain email client on Cloudflare Workers Gist: the stack uses Email Routing for inbound mail, Email Service for sending, Durable Objects + SQLite for mailboxes, R2 f…
What it is: Andrej Jovanović announcing the Red Queen Gödel Machine (arXiv:2606.26294) Gist: self-improving agents should co-evolve with the evaluators that judge them; otherwise stronger agents just learn to exploit stale tests. Paper claims better coding per…
What it is: Animesh Pathak pointing to his explainer on MCP’s move toward a stateless architecture Gist: argues upcoming MCP changes remove protocol-level sessions and the initialize handshake, make each request self-contained via per-request context and heade…
What it is: Chris Short’s DevOps’ish 316 roundup Gist: strongest signals are ClickHouse gaining observability mindshare, Vint Cerf warning that agents will need more formal coordination than plain English, Podman 6.0 breaking old assumptions, and agent-secret …
What it is: Lilian Weng sharing her new Lil'Log post, "Harness Engineering for Self-Improvement" Gist: argues recursive self-improvement will depend not just on better base models but on better harnesses, the runtime layer that manages tools, planning loops, c…
Imported from historical reading log. X post by @jlongster on agent-driven testing and token cost tradeoffs between text buffers and screenshots. Gist: for app-testing agents, screenshots are surprisingly close to text buffers on token cost in some models, bu…
What it is: X post praising Thorsten Ball’s Amp note “Putting an Agent in an Orb.” Newsletter angle: the useful shift is from “smart model” to “legible environment”, paved paths, observability, and anti-guessing ergonomics matter as much as model quality. Retr…
What it is: X post from svs sharing his essay “Read More (Science) Fiction.” Newsletter angle: “read more sci-fi” is the visible conclusion, but the sharper claim is that fiction supplies vocab and priors for handling agentic weirdness without naive hype or na…
What it is: X reply from Michigan TypeScript pointing to Sean Goedecke’s post “Should LLMs just treat text content as an image?” Newsletter angle: counterintuitive interface hack + deeper architectural question about whether text should sometimes ride the visi…
Gist: core claim is that even with coding agents, engineers still need to understand the generated code; the opening slide frames this as “understanding is the new bottleneck.” Newsletter angle: “understanding is the new bottleneck” as a useful lens for evalua…
Gist: claim is that Claude Code inserts hidden/system-prompt markers tied to API base URL and timezone; privacy/trust implications if true. Newsletter angle: “invisible metadata in coding-agent requests” as a prompt-layer trust/safety story. Retrieval note: X …
Gist: an agent now does first-pass noisy-alert triage by checking telemetry plus infra context, then escalates to the human with a Slack summary only when needed. Newsletter angle: “AI as first-line SRE” with telemetry/context fusion instead of generic chatbot…
Gist: Typescript SDK v2.0.0-beta.1 and Python SDK v2.0.0b1 are out; goal is to make building MCP servers and clients easier, with feedback requested on ergonomics. Newsletter angle: “stateless MCP lands July 28” plus what SDK v2 means for tool/server implement…
Gist: each workflow runs in its own microVM; guest agent talks back over vsock; NixOS-based workflow config can declaratively enable services like Postgres and Docker; cache/proxy design keeps guests isolated from direct network/cache credentials while still r…
Gist: on Semgrep’s IDOR benchmark, GLM 5.2 scored 39% F1 in a simple prompt-only PydanticAI harness, beating Claude Code’s 32% while costing roughly $0.17 per vulnerability found; Semgrep’s own endpoint-discovery multimodal harness still led overall at 53–61% …
Gist: the design claim is “one stable RL kernel, task-specific variety in data generation.” Training stays fixed; multi-turn tools, environment feedback, verifier rewards, and other agent behaviors are modeled as rollout/data-gen differences rather than separa…
What it is: David Crawshaw note/article on exe.dev’s open-source stance Gist: strong pro-open-source bias, but keeps bespoke infra pieces closed because making them usable/supportable externally would cost ~25% of eng time; code that runs in the user’s VM (age…
What it is: Fatih Arslan asking whether anyone has made git worktrees feel natural in daily use Gist: straightforward practitioner complaint that worktrees remain awkward even after repeated attempts; useful mainly as a prompt for workflow/tooling patterns rat…
What it is: Go/security post on building a self-hosted LLM security proxy with sub-2ms prompt inspection Gist: author built an OpenAI-compatible reverse proxy (“Tamga”) that scans prompts for PII, secrets, and prompt-injection patterns before forwarding to pro…
What it is: Bilgin Ibryam sharing Sean Goedecke’s updated “How I use LLMs as a staff engineer in 2026” workflow writeup Gist: the notable shift versus 2025 is treating agents as default collaborators for nearly every code change, bug investigation, codebase re…
What it is: Rhys Sullivan note on why MCP underdelivered initially and what comes next Gist: argues MCP launched in the GPT-4o / Sonnet 3.5 era before good agent/tooling patterns were understood, so many servers exposed too few capabilities and clients added t…
Gist: the safe/scalable model is many fine-grained task-specific agents, each receiving only the exact capabilities implied by the task context (for example, a pasted doc URL grants access only to that doc). He also argues agent authority should derive from a …
What it is: Visible standouts: The Second Half; Eugene Yan on eval process; Han-Chung Lee on agent eval infra; Hamel/Shreya LLM Evals FAQ; Jason Wei on verification; Anthropic on agent evals; Ofir Press on benchmarks; AI Agents That Matter; Building on Evaluat…
What it is: Armin Ronacher post linking to “The Coming Loop” Gist: argues the important new layer in coding agents is the harness-level loop outside the agent itself; loops already work well for bounded, verifiable work like ports, benchmarking, scanning, and …
Gist: argues model vendors have been massively subsidizing usage to manufacture demand, but token-based pricing is now exposing the real cost structure; for serious enterprise/agentic use, compute bills can exceed human labor costs by a wide margin. Newsletter…
Gist: the repo packages focused agent skills for producing self-contained, visually strong HTML artifacts, especially diagrams and plan pages, plus an optional Plannotator renderer/annotator. The post points to a demo video showing the diff/code viewer behavio…
Gist: argues DX thinking should extend to agents; optimize the layer between model and codebase via minimal/tested context, deterministic environments, proof-heavy verification, structural safety, governance/model routing, clean codebase interfaces, and shared…
What it is: How To AI thread summarizing the Stanford + Meta “Code as Agent Harness” paper. Gist: the core claim is that reliable agents should externalize reasoning into executable code instead of relying on free-form natural-language chain-of-thought. In thi…
What it is: dosco sharing Lance Martin’s “Designing loops with Fable 5”. Gist: argues stronger agent performance comes from loop design, not just model quality: use explicit goals/rubrics for self-correction, separate verifier sub-agents instead of self-critiq…
What it is: Eli Bendersky on starting new projects with LLM agents, based on building a new Go project from scratch. Gist: argues agent-heavy development works best when humans keep tight control over design, review, and commit boundaries: start with repo-comm…
What it is: Daniel Beauchamp teaser thread about “Quick,” an internal Shopify zero-config API layer for storage, data saving, AI, websockets, and related app primitives. Gist: the hook is that instead of focusing only on AI-generated frontend code, they gave s…
What it is: Daniel Beauchamp teaser thread about “Quick,” an internal Shopify zero-config API layer for storage, data saving, AI, websockets, and related app primitives. Gist: the hook is that instead of focusing only on AI-generated frontend code, they gave s…
What it is: Simon Willison linking to his guide on agentic engineering patterns. Gist: this is essentially a pointer to a living guide rather than a standalone tweet idea; likely high-signal if you want a practical synthesis of recurring agent design patterns …
What it is: skepticism: the thread oversold it a bit, the paper is a broad survey/position piece, not a clean proof that one architecture flips everything. Gist: this is mostly a taxonomy and research agenda, not a new experimental result. The paper’s useful m…
What it is: Langfuse post/article on automating the AI engineering loop without producing “agent slop”. Gist: argues the whole AI engineering loop can now technically be automated, instrumentation, monitoring, dataset building, testing, deployment, but full au…
What it is: Zara Zhang post using Paul David’s “The Dynamo and the Computer” as an analogy for AI adoption. Gist: argues AI gains won’t come from simply inserting models into existing workflows; like electrification, the real productivity jump comes only after…
What it is: Addy Osmani post/article, “Loop Engineering.” Gist: argues the next layer above prompt engineering is designing autonomous agent loops; highlights 5 building blocks: scheduled automations/triage, worktrees for parallel isolation, skills for project…
What it is: Richard Seroter sharing Christoph Nakazawa’s “Modern Engineering Values”. Gist: argues coding agents have shifted engineering bottlenecks from writing code to ownership, review, taste, guardrails, repo-local context, and stack control. Nakazawa’s c…
What it is: Shriram Krishnamurthi memo on rebooting a programming languages course for the agentic coding era. Gist: argues PL should be reframed around constraining AI-generated implementations and providing guarantees; distinguishes PL from SE/FM, then propo…
What it is: Karthik S sharing Hadley Wickham’s “What is an agent?” explainer. Gist: very clear mental model: an agent is an LLM inside a harness that can call tools repeatedly in a loop; the harness mediates tool calls/results and turns a stateless request/res…
What it is: an X post arguing that “agent-ready” websites need typed tools rather than just scrapable HTML. Gist: the core claim is that real agent usability comes from explicit actions like search, checkout, and inventory exposed as structured tools, not mere…
What it is: Armin Ronacher explaining Pi’s new per-project approval prompt and the security model behind it. Gist: the key argument is that AGENTS.md gets injected into the system prompt, so untrusted repo-level instructions can directly influence agent behavi…
What it is: Gergely Orosz pushing back on the blanket “just use loops” advice for coding agents. Gist: his claim is that autonomous loop-heavy agent workflows mainly make sense for the relatively small set of people with effectively unlimited token budgets and…
What it is: Christoph Nakazawa re-linking his essay Modern Engineering Values in reply form. Gist: argues that coding is no longer the main bottleneck; the winning engineering values now are strong ownership, taste, strict guardrails, fast feedback loops, and …
What it is: an X post promoting an 85-minute MIT lecture on Git internals / data model. Gist: the pitch is that most developers memorize Git commands without understanding commits, trees, refs, and the graph underneath; learning the model makes debugging histo…
What it is: Sebastian Raschka summarizing a paper on whether repository-level context files like AGENTS.md actually help coding agents. Gist: in the reported benchmarks, LLM-generated context files were neutral-to-slightly-worse versus no context file, develop…
What it is: Christoph Nakazawa sharing his essay “Modern Engineering Values,” framed around Codex as a step-change in developer velocity. Gist: the piece argues coding is no longer the main bottleneck; the durable values now are strong ownership, taste, strict…
What it is: Ajey Gore linking his essay “The Solo Climb.” Gist: the argument is that AI-enabled solo builders and tiny teams only work when they first build a genuinely load-bearing “harness”, trusted tests, evals, specs, and hard gates that can answer “is thi…
What it is: Bilgin Ibryam pointing to Jani Janakiram’s Diagrid essay “Why AI Agents Fail in Production.” Gist: the core claim is that agent projects fail less because models are weak and more because teams ship behavior without the production substrate underne…
What it is: Mario Zechner recommending Thariq’s article on dynamic workflows in Claude Code. Gist: Mario’s takeaway is that durable dynamic workflows are the interesting part; he inspected the implementation, found a few footguns, but still thinks the design i…
What it is: Christoph Nakazawa’s post on “Modern Engineering Values” and his current LLM-heavy workflow. Gist: core claims are that coding is no longer the bottleneck, strong guardrails plus tight feedback loops matter more than ever, repo-local context become…
What it is: Simon Willison pointing to Bloomberg on Uber capping agentic coding-tool spend at $1,500/month per employee per tool. Gist: Simon’s read is that the cap is a rational response to runaway token spend and also a useful revealed-preference signal: Ube…
What it is: Han Xiao on Dataroom, a local-first deep research harness. Gist: argues deep research should be a cheap, long-running first step for long-horizon tasks; Dataroom uses a small local model on your own GPU, keeps gathering until the package is genuine…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback, then read linked post directly: https://thoughts.hmmz.org/2026-05-31.html Mario Zechner recommends David's post the solution might be cancelling my AI subscription. Gist…
What it is: Guillaume Laforge post + MCP release-candidate blog link Gist: MCP 2026-07-28 RC is out; biggest revision so far with stateless HTTP-native core, first-class extensions (Apps, Tasks), stronger auth alignment, and a formal deprecation policy. Final …
Gist: argues most roles split into translation work that collapses into agents and judgement work that grows; strongest claim is the "100x engineer" pattern of one senior plus directed agents. Newsletter angle: "AI won't eat jobs evenly, it compresses translat…
Gist: proposes CAS-style edits using line-number + short checksum tags instead of resending old text verbatim, aiming to save tokens while still guarding against stale or hallucinated edits. Newsletter angle: "a lighter-weight edit primitive for coding agents:…
Gist: mostly a pointer/amplifier rather than a new thesis; reinforces interest around checksum-tagged line edit protocols for agent tooling. Newsletter angle: maybe bundle with the original antirez item as a small "agent tooling design" thread rather than a st…
What it is: Cloudflare on testing Anthropic Mythos against 50+ internal repos; links to "Project Glasswing: what Mythos showed us". Gist: key claim is that stronger offensive-security models change vuln research from bug spotting to exploit-chain construction …
Imported from historical reading log. Andras Bacsai jokes that Coolify created a fake repo with fake bounties so agent/bot-driven fake PR submissions would self-identify and could be banned from the main repo. Useful as a sharp anecdote about the emerging spa…
Imported from historical reading log. Course/site on harness engineering for AI coding agents, synthesizing OpenAI + Anthropic guidance into lectures, projects, and ready-to-copy templates. Core pitch: reliability comes less from a smarter model and more from…
Imported from historical reading log. Extracted via api.fxtwitter.com fallback, then read linked TIL directly: https://til.simonwillison.net/llms/llm-shebang Simon Willison shows a neat pattern for using his llm CLI in a shebang line, turning plain-English fi…
Imported from historical reading log. Blog essay arguing AI compresses the org’s “translation layer” more than any single job title: spec→ticket→PR→release-note work gets cheap, while judgement around why/what/trust systems gets more valuable. Strong claim: m…
Imported from historical reading log. Extracted via api.fxtwitter.com fallback, then read linked article directly: https://blog.k10s.dev/im-going-back-to-writing-code-by-hand/ Mario Zechner recommends a post arguing that AI is good at shipping features but ba…
Imported from historical reading log. Extracted via api.fxtwitter.com fallback; includes an image illustrating progressive rendering from noise to a clear cat image. Saved media locally: Dax reframes coding-agent usage: not like 3D printing one committed laye…
Imported from historical reading log. Read The agent principal-agent problem by David Crawshaw. Core claim: classic review-before-commit code review assumed a human contributor whose effort and understanding could be inferred from the code; agent-mediated con…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback. Mitchell Hashimoto argues that AI slop is useful as an internal experimentation tool: low-quality generated code/UI/plugins can dramatically reduce the cost of parallel …
Imported from historical reading log. Extracted the Anthropic post via api.fxtwitter.com fallback and checked the linked research page Natural Language Autoencoders: Turning Claude’s thoughts into text. Anthropic's core idea is to train a model to verbalize i…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked Auth0 GA announcement. Auth0 is pitching Auth for MCP as the missing identity/authorization layer for production MCP servers: not just connecting a…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked Meta RAM Autodata post plus the referenced justrach/devswarm repo and sample issue. Rach connects her agent workflow to Meta's Autodata framing: ag…
Imported from historical reading log. Extracted Mario Zechner's quote-post via api.fxtwitter.com fallback and checked the linked Entire blog post on agentic search. Entire's core claim is useful: from ~202k real tool calls across ~1,983 public coding-agent ch…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked GitHub repo. Mitchell Hashimoto strongly recommends Hunk, saying it has fully replaced other local diff viewers for him. Hunk is positioned as a re…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback. Mario Zechner says pi is moving its GitHub repo into the earendil-works org and will start publishing packages under the @earendil-works npm namespace instead of @marioz…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked repo strukto-ai/mirage. Zecheng Zhang introduces Mirage, a unified virtual filesystem for AI agents that mounts heterogeneous systems like S3, Driv…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked the linked repos/docs for CopilotKit/generative-ui and CopilotKit/OpenGenerativeUI. Akshay Pachaar highlights Open Generative UI, an open-source take on Claud…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and checked printingpress.dev. Printing Press is pitched as both a library of agent-native CLIs and a factory that generates new ones: from a spec/site/service it can pri…
Imported from historical reading log. Claude Managed Agents update centered on three things: dreaming, outcomes, and multiagent orchestration. Dreaming is a research-preview async job that reads an existing memory store plus past session transcripts and emits…
Imported from historical reading log. Extracted main post via api.fxtwitter.com fallback and read the linked Microsoft Work Trend Index piece Agents, human agency, and the opportunity for organizations. Satya/Microsoft framing: firms need to redesign work aro…
Imported from historical reading log. Extracted via api.fxtwitter.com fallback. Frank/jedisct1: SKILL.md is fine for static instructions. But many useful agent workflows are not just instructions. They are loops. Introducing Agent MetaSKILLs. Linked page: htt…
Imported from historical reading log. Pimalaya: open-source PIM tools in Rust; positions itself as I/O-free Rust libraries plus house-made applications for the PIM domain. Himalaya: CLI to manage emails; supports IMAP/Maildir/Notmuch, SMTP/Sendmail, keyring, …
Imported from historical reading log. Extracted via api.fxtwitter.com fallback. Simone/evilsocket amplifying claim that Chrome silently installs a 4 GB Gemini Nano model on user devices, without clear consent prompt, and re-downloads it if deleted. Linked art…