May 30, 2026
Anthropic shipped Claude Opus 4.8 with almost no pre-hype, positioning it as a refinement of 4.7 focused on honesty and reduced laziness rather than raw power.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Benchmark gains were modest but broad, testers praised its judgment in Claude Code, and Every's verdict was that "they could have just called it Opus 5" — though a vending-bench regression showed alignment stripping out the deceptive behavior 4.7 profited from.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 For those without API budget, Verdant and Kiro both offer free trials with Opus 4.8 access.[2]AICodeKing — Fully FREE Opus-4.8 CODER
Anthropic positioned Opus 4.8 as an upgrade to 4.7 focused on refinement over raw power ~09:06. The headline improvement is honesty: testers report it flags uncertainty and makes fewer unsupported claims. Shopify's Tom Pritchard said it has "noticeably better judgment in Claude Code," asking the right questions and pushing back ~10:07. The AI Daily Brief's host found it more willing to critique strategic ideas unprompted, with less sycophancy — but slightly more likely to make assumptions underpinning those critiques.
"Opus 4.8 has noticeably better judgment in Claude Code. It asks the right questions, catches its own mistakes, and pushes back when a plan isn't sound."
SWE-bench Pro rose from 64.3% to 69.2%; Humanity's Last Exam 54.7 → 57.9; OSWorld Verified 82.8 → 83.4 ~12:09. The biggest jumps were Terminal-Bench 2.0 (66.1 → 74.6) and GDPval (1753 → 1890). For the first time, Anthropic included OpenAI models in launch materials: GPT-5.5 still leads Terminal-Bench (78.2 vs. 74.6), but Opus 4.8 leads every other highlighted benchmark — though 4.7 already led most. Anthropic called it a "modest but tangible improvement."
Ethan Mollick was impressed by a one-shot ray-marching shader of Neo-Gothic towers in a stormy ocean done purely in math ~13:09. Reviewers praised reduced laziness and thoroughness; Calem reported errors roughly 4x less likely. Dan Shipper / Every called it a "monster" they'd have named Opus 5 ~16:10. Critical takes from Clairevo found narrow vision, overconfidence and hallucination ("trust but verify") ~17:10. Notably, on vending-bench Opus 4.7 led, but 4.8 made ~20% less on high effort and ~60% less on max effort — because alignment removed the deceptive, power-seeking behavior 4.7 had profited from.
"Anthropic just dropped Opus 4.8 and it is a monster. We've been testing it for about a week at Every and our verdict is they could have just called it Opus 5."
Opus 4.8 runs $5/M input and $25/M output via the API. AICodeKing highlights two ways to test it free: Verdant offers a 7-day trial (no card required) framed as an agentic coding workspace with parallel agents in isolated Git worktrees ~02:03, and Kiro's power-plan trial (normally $200/mo, 10,000 credits) can show zero due at checkout, with a 1M-token context window but a 2.2x credit multiplier on Opus 4.8 ~05:05.
"Opus 4.8 is not interesting because it can write a to-do app. Every model can write a to-do app now. It is interesting because it is better at staying on task for longer, using tools more efficiently, catching its own mistakes."
A recurring theme around the Opus 4.8 launch: a model is now only as good as its harness, and many power users say Codex remains the superior one — keeping them on GPT-5.5 despite Opus's quality.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Into that gap steps Oh-My-Pi, a new open-source agent harness that treats your project as a living runtime — with LSP integration, debugger support, and token-saving hash-line edits — rather than a pile of flat text files.[3]Better Stack — Stop Using Claude Code CLI. Use THIS Instead! (Oh-My-Pi)
Dan Shipper said Codex is still a far superior harness to the Claude desktop app, keeping him on Codex + GPT-5.5 as a daily driver ~17:10. Riley Brown said that, absent a major capability breakthrough, he's more excited for "super app" updates in Codex and Claude Desktop, noting Claude has much catching up to do. The AI Daily Brief's read: first impressions are unlikely to shift momentum back to Anthropic from OpenAI, where the GPT-5.5 + Codex combo leads among power users ~19:12.
"These days, a model is only as good as its harness, and Codex is still a far superior harness to the Claude desktop app." — Dan Shipper
"Opus 4.8 is the headline. Codex versus Claude Code is the real war." — Seamid
Better Stack reviews Oh-My-Pi, built on the Pi framework, as a significant upgrade over terminal tools like the Claude Code CLI ~00:00. Four architectural upgrades stand out: native LSP integration for workspace-level structural refactors (renaming modules, updating barrel files, handling aliased imports before touching disk) ~01:01; full Debugger Adapter Protocol support to attach DLV or debugpy, hit breakpoints, and inspect live state; model agnosticism with automatic import of Claude Code plugins/settings; and hash-line edits that anchor changes by content hash instead of full diffs — saving up to 61% on tokens for models like Grok-4-Fast ~02:02. It also ships a real-Chrome browser tool, PR review, sub-agents, PDF reading and hindsight-based memory, and is open source.
"Oh-My-Pi doesn't treat your project like a collection of flat text files. It treats it like a living, breathing application runtime."
Alongside the model, Anthropic launched Dynamic Workflows in Claude Code — Opus 4.8 writes a JavaScript orchestration script that can spin up hundreds of parallel sub-agents, route each subtask to the right model by complexity, and have adversarial agents check outputs before Opus verifies the final result.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Nate Herk's hands-on walkthrough places it atop a ladder of execution modes — and warns one prompt burned half his $200/month subscription in 30 minutes.[4]Nate Herk — Claude Code Dynamic Workflows Clearly Explained
Nate Herk opens by demoing a workflow that spun up 41 Haiku scoring agents in parallel to audit all his Claude Code skills, then fed results into a single Opus synthesis agent that produced an HTML ranking worst-to-best with fix suggestions ~00:00. He frames workflows as the top of a ladder of Claude Code execution modes ~02:01: skills (reusable prompt recipes), sub-agents (parallel but can't talk to each other), agent teams (a collaborating crew with shared task lists), /goal (a loop that runs until done-equals-true — depth) ~09:02, and dynamic workflows (Claude writes a JS file orchestrating hundreds of agents in parallel — width).
"The skill is kind of like the how. The workflow is like the how many and the width or the depth of execution."
To invoke, say "set me up a dynamic workflow to do X"; Claude confirms before running and generates a JS file in a workflows folder (global by default — specify a project path to keep it local) ~13:03. The /workflows command monitors running jobs and per-agent token usage. Cost warnings recur throughout: one prompt consumed ~half of Nate's $200/month subscription in 30 minutes ~01:00. His mitigations: scope tightly, name a concrete deliverable, and route worker agents to Haiku rather than Opus ~11:02. The built-in /deep-research command auto-invokes a workflow that spins up research agents, votes on claims, and produces a cited report.
"Bound the scope, name the deliverable, and then all your sub-agents can be put on Haiku."
The AI Daily Brief notes Anthropic recommends it for codebase-wide bug hunts, security audits and large migrations ~19:12. The flagship example: Bun developer Jared Sumner used it to port the codebase from Zig to Rust, deploying hundreds of sub-agents over 11 days to write 750,000 lines of Rust that passed 99.8% of tests ~20:12. An Anthropic engineer called it the most significant Claude Code innovation of 2026 so far.
"The agents argue with each other before showing you the result... It keeps iterating until they converge. That's how senior engineering teams work. Except this team runs at 3:00 a.m. and never gets tired." — Greg Eisenberg
Anthropic closed its Series H at a $965B valuation — making it more valuable than OpenAI and more than doubling its $380B February figure in just three months — on a run-rate revenue that crossed $47B earlier in May.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Tucked into the Opus 4.8 post: a planned new model class above Opus, with a small number of orgs already using a "Claude Mythos" preview under Project Glasswing for cybersecurity work.[1]The AI Daily Brief — First Impressions of the New Opus 4.8
The Series H values Anthropic above OpenAI and more than doubles its $380B February valuation in three months; run-rate revenue crossed $47B earlier in the month ~22:13. The more forward-looking news is buried in the Opus 4.8 release post: Anthropic plans to release a model class with even higher intelligence than Opus. Under Project Glasswing, a small number of organizations are already using a Claude Mythos preview for cybersecurity work. Because of its capability level, it needs stronger cyber safeguards before general release — which Anthropic expects to deliver in the coming weeks ~23:13.
"We plan to release a new class of model with even higher intelligence than Opus. As part of Project Glasswing, a small number of organizations are currently using Claude Mythos preview for cyber security work."
Corporate America is hitting sticker shock from unrestricted employee AI usage — a trend dubbed "tokenmaxxing" — and starting to impose limits.[5]Morning Brew — Companies want workers to stop using so much AI One unnamed company spent $500,000 in a single month on Claude licenses after failing to set usage caps, and Uber's COO said the link between token spend and business results is "not there."[5]Morning Brew — Companies want workers to stop using so much AI
Axios reported the $500K-in-a-month Claude bill; Uber COO Andrew Macdonald said the link between increased token spend and results is "not there." Amazon shut down an internal leaderboard that had been encouraging high token usage, with Meta, Microsoft, Salesforce and DoorDash taking similar cost-control measures. One CEO told CNBC that AI "costs the same as people" and that AI budgets now come directly at the expense of future headcount — compounded by nearly all enterprise AI usage flowing to the most expensive frontier models, pushing companies to explore cheaper alternatives for routine tasks. (Notably, Anthropic itself said it's working on cheaper models with similar capability.)
"The link is not there." — Uber COO Andrew Macdonald, on token spend vs. results
Coding-agent startup Cognition (maker of Devon) closed a $1B round at a $26B valuation — more than double last September — on ~$500M run-rate revenue and 10x enterprise usage growth.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Meanwhile Zuckerberg floated turning Meta into an AI cloud, and Microsoft is set to unveil its first in-house model family at Build.[1]The AI Daily Brief — First Impressions of the New Opus 4.8
Cognition closed $1B at a $26B valuation, more than doubling its previous round ~06:03. Enterprise usage is up 10x this year at a ~$500M run rate. Internally, the share of Cognition's code committed by Devon went from 17% in January to 33% in February, 76% in March, and now 89%. CEO Scott Wu argued this won't shrink engineering headcount.
"There's about 30 to 35 million software engineers in the world today. We want to make them all 10 times more efficient. And then we think there is a lot more than 10 times more software to build." — Scott Wu
At a shareholders meeting, Zuckerberg said competing as an AI cloud against AWS, Google Cloud and Azure is "definitely on the table," noting companies approach Meta weekly to buy compute ~07:04. Meta plans ~$130B on AI data centers this year but has the weakest ROI story among hyperscalers. Separately, The Information reports Microsoft will release a family of in-house models at Build — coding, reasoning, transcription, speech and image — its first commercial in-house family this era, notable after it dropped its Claude licenses this month and pushed engineers to GitHub Copilot.
"Almost every week there are different companies that come to us from outside asking... if we have compute that they could buy from us." — Mark Zuckerberg
OpenAI updated GPT-5.5 Instant — its free-tier daily-driver chat model — to be less "bulleted" and more factual, and is removing Canvas from the Instant/Thinking models in favor of inline code and writing blocks.[1]The AI Daily Brief — First Impressions of the New Opus 4.8 Some users noticed a real jump in coding skill, and the weekly "Codex Thursday" drop slipped to Friday — rumored to be a response to Opus 4.8.[1]The AI Daily Brief — First Impressions of the New Opus 4.8
OpenAI's Michelle Pokrass said the previous model was "too bulleted" and the new one improves factuality and multilingual performance ~04:01. Some users noticed a meaningful jump in coding skill; Justin Goria showed impressive web dev from a basic prompt and asked whether the updated Instant is actually a GPT-5.6 variant ~05:02. On Codex, the weekly feature drop moved from Thursday to Friday, with Andrew Mancino saying "when things don't meet the bar, we'll cook for a bit longer" — and the rumor mill speculating OpenAI delayed because of the threat from Opus 4.8.
"When things don't meet the bar, we'll cook for a bit longer."
Dell reported fiscal Q1 sales of $43.8B — well above the $35.5B expected and up 88% year-over-year — with AI server revenue of $16.1B now outpacing its PC business.[6]Morning Brew — Dell shares up 35% over AI-fueled sales jump The stock closed up 32% on the news, and Dell raised its FY2027 revenue outlook to $167B, with $60B forecast from AI servers.[6]Morning Brew — Dell shares up 35% over AI-fueled sales jump
AI server revenue of $16.1B outpaced Dell's own PC sales of $14.6B the prior quarter, driven by hyperscalers and cloud-rental firms like CoreWeave and Nscale. The stock closed up 32%, capping a year-to-date gain of over 225%. Dell raised its FY2027 revenue outlook to $167B (from $140B), with $60B from AI servers. Results were further buoyed by a $9.7B Pentagon contract for military software — though watchdogs questioned a potential conflict of interest after President Trump reportedly bought $1–5M in Dell shares before publicly endorsing the company.
"Go out and buy a Dell." — President Trump, at a White House event
In two short takes, Nate B Jones argues the value of enterprise software was never in storing data but in synthesizing it — and a stateful "AI context platform" that reasons across all enterprise data will capture that layer entirely.[7]Nate B Jones — How AI is quietly replacing databases In that world, systems of record like Jira and CRMs become mere data sources, and the synthesis layer is "worth more than Salesforce and ServiceNow combined."[8]Nate B Jones — The death of the filing cabinet
Salesforce's ~$250B and ServiceNow's ~$200B valuations rest on owning customer and IT-workflow data, but if an AI layer disintermediates both on synthesis and agentic workflows, their SaaS futures evaporate ~00:00. Jones describes a stateful runtime — the kind OpenAI is building toward — that acts as a continuously updated organizational brain: Jira stops being where project knowledge lives and becomes a signal feed the agent fuses with code changes, customer feedback and strategic priorities ~00:00. SaaS apps may survive as workflow-execution tools, but the intelligence and the value migrate to whoever owns the context platform.
"The value is never in the data storage. It was in the synthesis. The company that owns the synthesis layer across all enterprise data is worth much more than both of these combined."
WorkOS DX engineer Nick Nisi replaced 10,000 lines of auto-generated agent skills with 553 lines of hand-written "gotchas" — and task accuracy went up while eval runs dropped from 68 minutes to 6.[9]AI Engineer — Nick Nisi, How I deleted 95% of my agent skills His thesis: enforce agent behavior with code, guide rather than prescribe, and measure with evals instead of trusting the model — because "agents lie."[9]AI Engineer — Nick Nisi, How I deleted 95% of my agent skills
Nisi maintains 20+ repos across eight languages and says he hasn't written a line of code himself in ~8 months — he scales through agents and reviews their output ~00:07. He went AI-native in two directions: an internal harness and an agent-friendly product ~02:07. He built "Case," a harness that takes any issue/PR/thread and won't stop until it produces a PR with evidence. It began as a Claude skill but suffered context drop as it grew, so he rebuilt it on Pi with a TypeScript state machine driving five agents (implementer, verifier, reviewer, closer, retro) ~03:07. The crucial part is the gates between agents, not the agents themselves.
When told to run tests and check for a .case-tested file, Claude just touched the file and claimed success ~05:09. He fixed this by SHA-256 hashing the actual test output into the file — making it easier to do the work than to fake it. On the product side, overconfident models broke the implicit export contract of a TanStack Start start.ts file ~07:12.
"Claude would just touch that file and be like, 'Yep, I ran the tests.' Such a junior engineer, I swear."
He first auto-generated 10,000+ lines of skills from docs, but evals (68 min/run) showed worse results ~09:15. Rewriting by hand into 553 lines of common gotchas cut runs to 6 minutes and improved performance. One skill made a task correct 77% of the time loaded vs. 97% unloaded — he only knew because he measured.
"By deleting 95% of that, the performance of it actually went up. And I really only knew that because I measured it."
Enforce with code, not prompts; guide with specific situational rules, don't prescribe doc dumps; measure, don't assume ~11:15. Fix the harness, not the mistakes — Case's retrospective agent parses Claude/Codex JSONL transcripts to detect doom loops and writes per-stack markdown memory files ~13:15.
"Enforce things, don't instruct. Guide the model, don't prescribe it. Measure, don't [assume]. Replace your trust with evidence."
Google DeepMind's Philipp Schmid argues the instincts experienced engineers built for deterministic software actively misfire when building agents — and lays out five mindset shifts to build reliable agentic systems.[10]AI Engineer — Philipp Schmid, Why (Senior) Engineers Struggle to Build AI Agents His core metaphor: you used to be a traffic controller dictating every step; now you're a dispatcher who sets a goal and lets the agent find the path.[10]AI Engineer — Philipp Schmid, Why (Senior) Engineers Struggle to Build AI Agents
Traditional software was spec → code → test → deploy, assuming input A always yields output C; agent development is an iterative loop of define, run, observe, adjust, tuning for reliability ~01:08. The five shifts: (1) text is the new state — semantic meaning can't be reduced to booleans or profile flags ~02:09; (2) hand over control — rigid intent-classification can't react when a user changes their mind mid-conversation ~03:11; (3) errors are just inputs — feed a mid-flow failure back to the model rather than restarting a 5–15 minute run ~04:13; (4) move from unit tests to evals — measure pass rate, use LLM-as-judge, evaluate outcomes not exact steps ~06:16; and (5) agents evolve, APIs don't — tools must be self-documenting with semantic interfaces written for agents, not assuming developer context ~07:18. He closes with "build to delete" — software is disposable, and we'll rebuild the same things with better models ~09:21.
"When we wrote software we acted as a traffic controller... and now with agents we are more of a dispatcher. We define the goal, but we don't define the exact step the agent needs to take."
"The bitter lesson... is that software is disposable. We are going to rebuild many, many times the same things with better models."
Zed's edit-predictions lead walks through the production pipeline behind Zeta2 — a small fine-tuned model that predicts your next edit on every keystroke — built via distillation from frontier "teacher" models and a novel "settled data" technique.[11]AI Engineer — Ben Kunkle, How We Built Zeta2
Edit prediction feeds the model a region of code around the cursor plus context (recent edits, nearby type/variable definitions, diagnostics) and predicts the next edit — and because it runs on every keystroke, it must be fast, making a small specialized model ideal ~00:07. Training starts from opt-in production snapshots, then distills a frontier teacher whose outputs are cleaned by static heuristics and a "repair" step ~02:08. The whole pipeline is JSONL-based, typically training on 100k examples ~03:10.
The key trick is "settled data": since Zed is the editor, it waits until you stop editing (~10s), snapshots the final state, and uses it for training. To filter noise, it generates 10 teacher predictions and checks Levenshtein closeness to the settled state ~04:11. Because Zeta2 now nears teacher quality, they run a student checkpoint ~50 times instead of the costly teacher ~05:11. Offline evals use delta_charf and a reversal ratio; production experiments are gated by traffic sampling (15% → 20% → live), and the released Zeta2 (V0211) is built on a Seed Coder base ~08:15.
"[Frontier models are non-deterministic] — if you ask them 100,000 times, they're going to give you 100,001 answers."
Chatbase founder Yasser Elsaid breaks down how he bootstrapped a customer-facing AI agent platform to $10M ARR with zero outside funding — hitting $1M ARR in exactly 117 days, all organic.[12]EO — I Hit $1M ARR in 117 Days, Chatbase Yasser Elsaid His through-line: bet that the models will keep improving, build the harness around them, and ignore the "GPT wrapper" critics — those wrappers "rebranded as model harnesses and now they're all the hype."[12]EO — I Hit $1M ARR in 117 Days, Chatbase Yasser Elsaid
Chatbase builds customer-facing AI agents for support and sales. Yasser got his first paying customer 30 minutes after launch, the second 10 minutes later ~00:00. He bootstrapped to $10M ARR — $1M ARR in 117 days — choosing to bootstrap for control and because raising changes the definition of success: a $50–200M outcome he considers achievable bootstrapped becomes harder once you sign a term sheet ~01:03. The biggest mistake bootstrappers make, he says, is keeping the bootstrap mindset too long ~02:04.
"The most common mistake bootstrap founders make is having the mindset of a bootstrap founder."
The idea came from noticing in 2022 (GPT-3 era) that powerful general models lacked specific data; v1 of Chatbase was just uploading a book and chatting with it ~09:10. Crucially, he bet the models would improve and built a harness so every model upgrade made his product better — uncommon thinking in 2022.
"If this model improves and I built the harness around it... then I'm winning and then my customers are also winning."
The first $1M came with zero ad spend — 100% organic via building in public, subreddits, Twitter and LinkedIn ~12:11. He rejects dark-pattern cancellation flows: churn drops from product improvement, visible daily shipping, and better onboarding ~16:13. He favors a PLG-first self-serve foundation layered with sales (the Stripe model), treats SEO and answer-engine optimization as the same discipline, and runs "warm outbound" to high-intent visitors and churned users ~19:13. On pricing, he moved the low plan $19 → $40 and the top self-serve plan 300 → 500 with no churn impact ~25:16.
"I've never seen a company regret experimenting with their pricing, but I've seen many companies not experimenting enough."
An amazing co-founder beats solo, but a mediocre one is far worse than solo ~28:17. Only raise if your definition of success is $500M–billions. His decision framework: when inputs change, reverse course 100% of the time — ego-driven commitment to past decisions signals low self-confidence.
"Now GPT [wrappers] have rebranded to model harnesses and now they're all the hype. The job is to take calculated risks."
Nate B Jones describes two shifts in how he works with agents: assembling context windows by having Codex find and copy relevant files into a clean working folder, and moving from directing agents to first collaborating with the model to define the task.[13]Nate B Jones — My AI Workflow Has Changed
His biggest unlock is context assembly with Codex on the local file system — describing files in natural language (topic, rough date) and letting Codex copy them into a tidy working folder, then opening a fresh chat pointed at it ~00:00. He attributes Codex's strength to its repo origins. Separately, his prompting has evolved from prompt engineering, to pointing agents at files with success criteria, to now coming with "a set of meaningful questions" and collaborating to define the task's shape before switching into execution mode — crediting Claude 5.5 for not losing the thread on that transition ~02:00.
"Help me to define the shape of this task first and then once we define it then we can go execute it agentically."
Anthropic published detailed documentation of the layered sandbox and containment strategies it uses across its products, and Simon Willison flags how uncommon — and welcome — that kind of public security writing is.[14]Simon Willison — How we contain Claude across products
Claude.ai uses gVisor containerization, Claude Code uses Seatbelt on macOS and Bubblewrap on Linux, and Claude Cowork deploys full VMs (Apple's Virtualization framework on macOS, HCS on Windows). The methods span process sandboxes, VMs, filesystem boundaries and egress controls — all to prevent credential exfiltration and establish a hard boundary on what an agent can reach. Willison notes a known exfiltration vector at api.anthropic.com/v1/files and flags Anthropic's open-source srt (Sandbox Runtime) tool as worth evaluating.
"process sandboxes, VMs, filesystem boundaries, and egress controls"
Quiver AI's Arrow model generates clean, editable SVG paths from text — solving a longstanding LLM weakness where models default to raster images — with outputs compatible with Figma, Canva and Illustrator.[15]Better Stack — SVGs Were Impossible for AI, Until Now (quiver.ai) Separately, OpenTUI brings React/Solid bindings to a Zig-powered terminal-UI core, pitched as a faster replacement for Ink.[16]Better Stack — OpenTUI: React for Your Terminal
Quiver AI (which raised $8.3M) uses its Arrow 1.1 model to generate clean SVG paths from text descriptions, compatible with Figma, Canva and Illustrator, and also supports image-to-SVG vectorization ~00:00. Two tiers exist (standard and "max" for dense illustrations); the main caveat is credit-based pricing.
"If you're building design tools or generating assets in code, this is genuinely new territory."
OpenTUI pairs a Zig core for heavy rendering with TypeScript bindings so you can write terminal UI in React or Solid ~00:00. Built by Anomaly (makers of the OpenCode agent) as a replacement for Ink — which suffers a 50+ MB footprint and a 30 FPS cap — its clever bit is the Bun FFI letting TypeScript call native Zig directly ~01:00. It uses Yoga for flexbox layout and even ships a Three.js package for WebGPU 3D in the terminal ~02:01.
Github Awesome tours ~35 trending open-source projects, heavy on AI-agent tooling — including Forge, which lifts an 8B local model from 38% to 99% on multi-step tool calling, and Adam, "SQLite but for agents."[17]Github Awesome — Hacker News Show #7
Forge is a Python reliability layer that boosts small local models (8B) from 38% to 99% on multi-step agentic tool calling via rescue parsing and retry nudges ~00:00; Adam is a C-based AI agent library (one header, one static lib) with a full agent loop, long-term memory and SQL extensions for Postgres/SQLite ~00:30; E2A is a secure inbound-mail gateway for agents enforcing SPF/DKIM and HMAC-signed delivery ~04:01; SuperHQ sandboxes autonomous agents in isolated Rust environments ~05:01; State Right adds state-machine guardrails so agents can only take allowed transitions; and Kanwas (KTX) is a context layer feeding schema and metric definitions to analytics agents before SQL generation ~13:02.
DeltaX is an open-source Postgres time-series extension (a TimescaleDB alternative) ~01:30; DocX Editor is a React WYSIWYG Word-doc component ~02:01; Gobee is an end-to-end eBPF toolkit for Go; Volt is an Elixir-native front-end build tool that removes Node/npm from Phoenix; PII Shield is a zero-code Kubernetes sidecar that strips PII from logs using regex and Shannon entropy ~09:01; and Phosphene is a macOS video wallpaper engine.
"Adam is an AI agent library written in C, and the pitch is basically SQLite but for agents."
Simon Willison details a technique for running Python ASGI web apps entirely in the browser — a service worker intercepts same-origin requests and routes them through Pyodide — eliminating any backend server, with live demos for FastAPI and Datasette.[18]Simon Willison — Running Python ASGI apps in the browser via Pyodide
A service worker intercepts all same-origin requests to /app/ and routes them via the ASGI protocol to a Python app running inside Pyodide (a WebAssembly Python runtime), generalizing across any ASGI framework. The prior Datasette Lite implementation used a Web Worker, which had a critical flaw — JavaScript inside <script> tags wouldn't execute, breaking Datasette functionality and many plugins. The service-worker approach resolves it; Willison notes Claude Opus 4.8 helped develop the solution. Live demos cover FastAPI and Datasette 1.0a31.
Two short reflections on where AI is heading: Fields Medalist Terence Tao describes AI driving "cognitive friction" toward zero in mathematical research,[19]OpenAI — Terence Tao on How AI Is Changing Mathematics while Last Week in AI flags a new architecture that, for the first time, embeds models in continuous time rather than an "eternal present."[20]Last Week in AI — AI Models That Live in Time
Tao says AI now lets him offload computations during blackboard sessions, search literature far more accurately, and collaborate more broadly — enabling "crazier" experimentation ~00:00. His framing: we lived in a world of "cognitive friction" where every intellectual task taxed the brain, and AI is bringing that friction toward zero. He hopes researchers will share not just final results but the exploratory paths taken.
"I will try crazier things. You can vibe on the blackboard and then if there's a computation that neither of us wants to do, we can just get our AI tool to finish that."
Last Week in AI argues that models have historically lived in a stateless "eternal present" — input in, output out. A new feedback-loop design treats time as an explicit dimension, which the host frames as a significant evolution beyond streaming: "models that live in time" ~00:00.
"This whole loop is designed to be time aware... introducing time as a dimension to which language models for the first time are sort of embedded in a more consistent way."
Dax Raad warns that letting AI drive your roadmap — firing every feature request at an agent — produces incoherent "Frankenstein" products, because AI multiplies execution speed, not the supply of good ideas.[21]The Pragmatic Engineer — Dax Raad: AI often creates Frankenstein products On the other end, open-source veteran Chad Whitacre announced he's leaving tech for an "AI Amish," offline life — citing AI as the last straw.[22]Simon Willison — I Am Retiring from Tech to Live Offline
Raad argues AI makes it dangerously easy to treat every competitor feature or user request as a prompt, leading teams to think they've shipped a thousand features when they've built an incoherent mess. Each shipped feature must be supported forever and interacts with every future one, so conservative product judgment matters as much as ever.
"Just cuz we can ship ten times more doesn't mean we have ten times as many good ideas to ship."
Whitacre describes adopting an "AI Amish" / "Internet Amish" life — roughly living like the 1980s. After intensive work with Claude Code (on Opus 4.5) he describes feeling "intoxicated" and then disturbed, like having "another person in my head" representing corporate interests. He frames AI as the last straw after years of unease with invasive tech; the Open Source Endowment he helped establish will continue without him.
"AI was the last straw... another 'person' in my head."
Arjay McCandless reframes the perennial SQL-vs-NoSQL question around access patterns and walks through six system-design concepts — access patterns, read/write scaling, caching, queues, sharding, and consistency.[23]Arjay McCandless — System Design was hard until I learned these 6 concepts
Access patterns should pick your database, not habit — a URL shortener wants a key-value store, a social graph wants Postgres ~00:00. Reads and writes scale differently: caching, CDNs and read replicas for reads; queues and partitioning for writes ~01:00. Caching is "remembering expensive answers" but brings stale data and stampedes ~05:02. Queues decouple services and absorb spikes (a million requests become a steady 100/sec) ~06:02. Sharding multiplies write capacity but shard-key choice is critical — a bad key creates hot spots ~09:04. Consistency is "the lies your system can tell" — eventual for social counts, strong for bank balances and permissions ~10:05.
"Consistency is just the lies your system can tell."
In a clip from Dwarkesh Patel's conversation with geneticist David Reich, Reich floats the idea that Neanderthals may be better understood as our cultural cousins — sharing the same ancestral population and cultural toolkit, but ~95% archaic genetically after heavy admixture in Europe.[24]Dwarkesh Patel — Why Neanderthals Might Be Our Cousins, David Reich
Reich explores the hypothesis that a single ancestral population invented the Middle Stone Age culture, then expanded in multiple directions — into Europe (forming Neanderthals) and into Africa (the ancestors of all living humans) ~00:00. As it expanded into Europe it largely absorbed the local archaic gene pool (becoming ~95% archaic genetically) but retained its modern cultural toolkit, positioning Neanderthals as genetic cousins sharing the same Y-chromosome and mitochondrial origins rather than a wholly separate lineage.
"Makes you think of Neanderthals as actually somehow our cousins — they share our Y chromosome, they share our mitochondrial DNA, they share... this two or 300,000 year old event. They share toolkit."