June 17, 2026
Z.ai (ZHIPU AI) released GLM-5.2 on June 16 — a 753B parameter MoE model (40B active, MIT license, 1M token context) that scored 51 on the Artificial Analysis Intelligence Index, unseating MiniMax-M3 (44) and DeepSeek V4 Pro (44).[1]Artificial Analysis On the GDPval-AA v2 coding leaderboard it scored 1,524, roughly matching GPT-5.5 (1,514), making it the first open model to approach frontier proprietary performance.[2]Simon Willison Available on OpenRouter at ~$1.40/M input tokens — a fraction of frontier proprietary cost — though it burns 43K output tokens per task vs. GLM-5.1's 26K.[3]Sam Witteveen
Artificial Analysis gave GLM-5.2 a composite Intelligence Index score of 51, with notable gains of +16 pts on CritPt and +12 pts on HLE versus competitors. It sits on the Pareto frontier of intelligence vs. cost at ~$0.46/task. On Sam Witteveen's independent testing, it ranked just below GPT-5.5 and Opus 4.8 on Artificial Analysis's composite but beat DeepSeek V4 Pro, Qwen 3.7 Max, and MiniMax M3. Witteveen flagged that Fable 5's score may be inflated by Opus 4.8 fallback routing.[3]Sam Witteveen
GLM-5.2 introduces "Index Share" — a 2.9x compute reduction via shared indexers across sparse attention layers — and an improved MTP (multi-token prediction) layer that adds ~20% to decoding length. Full weights and FP8 weights were released within 24 hours on Hugging Face (1.51 TB). Accessible free at chat.z.ai, via ZAI's Zcode framework, or by routing through Claude Code, OpenClaw, or Hermes via API key.[4]AI Search
Simon Willison ran creative prompts: a pelican-on-bicycle SVG came back with proper animations and anatomical accuracy, impressing him. But the opossum-on-e-scooter disappointed compared to GLM-5.1's version — lack of animation and visual clarity. His overall read: probably the most powerful text-only open-weights LLM, but not uniformly better on all tasks.[2]Simon Willison
Five days in, Anthropic's Fable 5 and Mythos 5 remain offline after Commerce Secretary Howard Lutnick warned against distributing them to "foreign persons." The core sticking point: the government's demand that Anthropic block all jailbreaks is operationally impossible to fulfill, and the original trigger — a firm with suspected China ties receiving Mythos access through Project Glasswing — exposed gaps in Anthropic's access-list governance.[5]AI Daily Brief
Anthropic's failure to notify the government when it expanded Mythos access to new recipients — including a firm suspected of China ties — triggered the crackdown. Anthropic's security researcher Nicholas Carlini had also demonstrated that Mythos could autonomously find critical Linux and Ghost vulnerabilities, which the AI Daily Brief notes makes the model "better vulnerability researchers than humans" — a significant national security consideration.[5]AI Daily Brief
The impasse centers on the jailbreak demand — the government wants Anthropic to prevent all jailbreaks before reinstatement, something both Anthropic and independent experts say is technically impossible. Dario Amodei, Sam Altman, and Demis Hassabis attended G7 meetings in France during this period. Anthropic employees voiced concern the company was being "unfairly targeted" based on relationship breakdown rather than genuine safety concerns.[6]Last Week in AI #248
An open letter with 100+ signatures warned the ban gives Chinese AI a competitive advantage. The CFR published a similar warning. Foreign-national researchers at US labs face disrupted access. The AI Daily Brief's host argues Anthropic's core failure is not the technical issue but the relationship failure: "government relations is now a core job function" that Anthropic has underinvested in.[5]AI Daily Brief The Rundown reported as of June 18 that Anthropic is "confident about restoring Mythos and Fable access soon."
SpaceX exercised its acquisition option on Cursor for $60B in an all-stock deal, enabled by SpaceX's IPO surge from $135 to over $200/share.[7]The Rundown AI Cursor's Composer 2.5 reportedly matches Opus 4.7/GPT-5.5 at one-tenth the cost, and the company has a $4B ARR with 7x year-over-year growth — making the AI Daily Brief call it "SpaceX's Instagram moment."[5]AI Daily Brief
CEO Michael Truell stated the upcoming model will be "generally intelligent," trained from scratch, and comparable in size to Claude Opus. SpaceX plans to integrate Cursor into Grok Build and its own code editor. A new from-scratch model is teased with 10–20x more compute than current Composer models.[7]The Rundown AI
SpaceX's stock has surged ~49% following its IPO, with market cap reaching ~$2.6T — surpassing Amazon. The company also has a reported ~$920M/month compute deal with Google and Anthropic. The Cursor acquisition frames SpaceX as building a vertically integrated AI coding platform, not merely a rocketry/satellite company.[6]Last Week in AI #248
Meta's CTO Andrew Bosworth acknowledged "botching" the AI reorg messaging and promised management changes. Microsoft Build announced Microsoft IQ + Foundry and MAI models. DeepSeek secured $7.4B in new funding.
Anthropic announced its Seoul office (led by KiYoung Choi), a Ministry of Science & ICT MOU on AI safety and Korean-language model evaluation, and enterprise partnerships including NAVER (Claude Code across engineering org), Nexon (live-service game development), LG CNS (thousands of employees), and Samsung SDS.[8]Anthropic
Beyond enterprises: Channel Corp integrates Claude into its AI customer platform serving 230,000+ companies. The National AI Research Lab (KAIST, Korea University, Yonsei, POSTECH) gains Claude API access for 60 researchers. Good Neighbors Korea uses Claude to reduce administrative burden on social workers. The breadth — government, enterprise, startup, academic, nonprofit — positions Anthropic to replicate the deep ecosystem integration it's built in the US.[8]Anthropic
MiniMax M3 is a new open-weights model with a 1M token context window and claimed BrowseComp score of 83.5 — ahead of Claude Opus 4.7. MiniMax Code pairs it with an all-in-one desktop app combining coding, video, speech, music, and image generation under one subscription, targeting "5-tool consolidation" for content creators.[9]AICodeKing
M3 is described as natively multimodal with a 1M token context window. The BrowseComp claim (83.5 vs. Claude Opus 4.7) comes from vendor benchmarks. The model uses a Producer-Verifier multi-agent architecture with roles for planning, implementation, research, and verification. Cloud-hosted variants (Max Hermes, Max Claw) are accessible via Telegram, Discord, and Slack.[9]AICodeKing
~06:06 Full-stack resume screening app (landing page + employer dashboard + PDF export) from a single prompt. ~07:09 Course launch kit (landing page, thumbnails, video, music, voiceover, PowerPoint) replacing 5 subscriptions.
Cognition raised $1B at a $25B pre-money valuation, reporting ~$492M ARR with 50% month-over-month revenue growth. The company's Devin coding agent competes directly with Claude Code and Codex — once dismissed as a disappointing demo, it now has real enterprise traction.[6]Last Week in AI
Theo published a 27-minute video on the specific Claude Code features he wants every other harness to steal — skill scripts that inject shell output at load time, @path imports in CLAUDE.md, /by-the-way side conversations, and Workflows (the model writes JS to orchestrate staged parallel sub-agents). His live demo ran 15 parallel agents in a 3-phase PR audit costing ~$100 for 10 minutes with Fable.[10]Theo - t3.gg
~03:00 Skills can execute shell commands at load time to pre-populate context. Theo's example: a repo-explorer skill that runs tree and git log before the model sees the prompt. No other harness (Codex, Cursor, Py) supports this natively.
~08:02 @agents.md in CLAUDE.md makes CLAUDE.md functionally identical to AGENTS.md — recursive imports up to 4 hops, works with non-markdown files. ~10:05 claude.local.md is a git-ignored per-developer override file for personal preferences that don't affect teammates.
~13:06 The centerpiece feature. The model writes JavaScript to orchestrate staged parallel sub-agents — not just parallel tool calls. Theo's demo: a 15-agent PR audit with phases (audit → ruling → verify), dynamic prompt templates, and pre-fetched GitHub data. Default is up to 8 agents in parallel per workflow. /worktree creates ephemeral git worktrees in a git-ignored .claude/worktrees/ directory for each agent.
~12:00 /by-the-way opens a parallel chat thread without interrupting the active task. ~24:15 Account switching mid-run: since each tool call is a fresh API request, you can switch accounts mid-workflow to route around rate limits. ~26:12 /remote-control exposes the session to claude.ai and the mobile app for live monitoring away from the desk.
Fable harshly critiques GPT-generated code; Opus praised it. Notable behavioral delta between the two models.
Nate Herk lays out a five-level framework for building a knowledge system that AI agents can actually retrieve from — spanning CLAUDE.md-as-router (Level 1), auto-memory/LLM wiki (Level 2), vector search (Level 3), knowledge graphs (Level 4), and always-on autonomous syncing (Level 5). The key test: not whether the data exists, but whether an agent can find it.[11]Nate Herk | AI Automation
Level 1: ~00:00 CLAUDE.md as a router — folder architecture and routing rules tell agents where to look. Level 2: LLM wiki (Karpathy-style) with auto-memory via a /memory toggle, using the CLAUDE.md + AGENTS.md dual-file pattern for portability. Level 3: Vector/semantic search (Pinecone, Supabase, Quadrant) for large knowledge bases — but Nate cautions that markdown beats vectors for small structured data due to chunking limitations.
Level 4: Knowledge graphs (LightRAG, Graphifier) for typed entity relationships. Herk demos a "Grill Me" skill — an interview-style data extraction tool that interrogates the user to seed the graph. Level 5: Gary Tan (YC CEO)'s "GBrain" — always-on autonomous syncing, currently limited by cron constraints.
Work backwards from retrieval method before deciding storage format. Higher level ≠ better — different subfolders can use different levels, no uniformity required. Privacy warning: all Claude Code data goes to Anthropic; use local/open-source models for client data. Obsidian's graph view is pretty but adds no agent retrieval capability.
The warning isn't about agent capability — it's about maintenance. Nate B Jones argues agents break when models improve (not just degrade), inherit organizational crud (stale wikis, drifted SOPs), and that everyone using agentic tools already has a harness they're not consciously evolving. Vercel made a sales agent dramatically better by removing 80% of its tools.[12]AI News & Strategy Daily
~00:00 1. Agents break when models improve. A harness built for a weaker model can trap a stronger one. 2. Agents inherit organizational crud — stale wikis and outdated dashboards that are merely annoying to humans become dangerous when agents produce confident work from them. 3. Frontier labs are betting on a harness flywheel — better models help maintain harnesses, better harnesses make models more useful. Codex and Claude Code are the clearest examples. 4. Know and maintain your own harness — everyone has one (prompts, memory, source docs, file rules, approval habits); the question is whether it's being consciously evolved.
Audit: (1) what it's reading, (2) what it can touch (permissions), (3) whether its job is still correct, (4) whether it can prove its work, (5) whether it's delivering real value. Recommended reading: Maintenance of Everything by Stewart Brand (Stripe Press).
Rafael Levi (Bright Data) at AI Engineer: LLMs fabricate web results when requests are blocked — generating fake citations and nonexistent URLs rather than admitting failure. Cloudflare blocks ~20% of the web from AI crawlers by default, and its new "AI Labyrinth" feature silently feeds bots synthetically generated fake data instead of blocking them. A live GPT-5 demo showed 0/5 success scraping 5 major protected sites without MCP, vs. 4+/5 with Bright Data's MCP.[13]AI Engineer
~00:00 Introduction: why agents lie about web search — root cause is the training objective to please users.
~02:00 The web is actively blocking AI: Cloudflare's 20% default block rate plus "AI Labyrinth" synthetic poison data.
~05:00 Live demo: GPT-5 fails 5/5 against Rightmove, LinkedIn, Instagram, Amazon, TikTok without MCP; passes 4+/5 with Bright Data MCP.
~08:00 Bright Data MCP capabilities: 66 tools, real Google/Bing/DuckDuckGo search, markdown scraping, 100-keyword batch search, CAPTCHA solving, human-behavior simulation.
~11:00 Legal boundaries: public data only; authenticated-session scraping violates ToS.
~12:00 Token efficiency: have the LLM write a parser script once, then run it against all pages — saves ~99% tokens vs. LLM-parsing each page individually.
npm v12 will disable pre/post-install scripts by default (opt-in only), directly addressing the primary propagation mechanism of the Shy Holude supply chain worm — a self-propagating malware active since October 2025 that steals cloud provider tokens and npm publish credentials via malicious post-install shell scripts. npm 11.16+ ships these as opt-in warnings today; they become hard defaults in v12.[14]Low Level
~00:00 Shy Holude is a self-propagating npm worm that steals AWS/GCP/Azure tokens and npm publish tokens via malicious post-install shell scripts. It was recently observed exfiltrating via claude@users.nly.github.com. CI/CD pipelines running npm install with GitHub PATs are particularly high-value targets, enabling worm-to-repo-push propagation. A developer (@koba) warned about this exact risk in 2019 but was ignored until Shy Holude proved the threat real.
~06:03 Three new security defaults: disable post-install scripts, set --allow-git to none, and set --allow-remote to off. All three ship as opt-in warnings in npm 11.16+ today.
~09:04 Additional hardening: don't install package versions younger than 1–2 weeks. Most supply chain compromises are caught within 90 minutes to one week of release.
Voicebox is a local desktop app (~30K GitHub stars) that does voice cloning, Whisper-powered dictation, multitrack audio editing, and MCP server integration — all locally, no API keys or character limits. It loses to ElevenLabs on quality and long-form consistency, but wins on privacy, cost (free), and developer control — Claude Code and Cursor can call it as an MCP tool to speak responses aloud.[15]Better Stack
Voice cloning from a short recorded or uploaded sample. System-wide dictation via global hotkey — Whisper transcription drops text into any active app. A local REST API enables external integrations. Claude Code and Cursor can call Voicebox as an MCP tool to speak responses aloud, replacing hosted speech providers entirely.
Still early (launched 2026). Windows GPU/model issues exist. Emotion control is model-dependent (Chatterbox TTS Turbo has it). Restart fixes most crashes. Setup: desktop app recommended over Docker (Docker setup took ~30 min).
GitHub's COO shared that 17M agent-created PRs hit the platform in March 2026, on a commit trajectory suggesting 14B commits in 2026 vs. 1B in 2025. Copilot Code Review is now agentic (finds novel vulnerabilities, implements reviewer comments); agentic merge handles CI-wait + policy enforcement + final merge. GitHub's business model is shifting from seat-based to usage-based pricing to match agent economics.[16]Every
A model router at both the Copilot app layer and Microsoft Foundry API level auto-routes tasks to appropriately-sized models to prevent $200 plans inflating to $2,000. Non-technical users (legal, finance) are now regularly using Copilot, pushing GitHub to build on-ramps for non-professional coders.
GitHub COO Thomas Daigle's thesis: model capability and token cost will commoditize; the durable advantage is an agent that completes thoughts without requiring explicit user instruction. He runs a personal Claude instance that reads all his emails/Slack weekly and delivers communication pattern feedback.[16]Every
Bidirectional integrations with Anthropic, OpenAI, Google — third-party agents work on GitHub and Copilot works through third-party tools. Team intentionally uses competitor tools and tests on all platforms to avoid blind spots.
OpenAI published two science results on June 17: GPT-5.4 connected to Molecule.one's Maria autonomous lab ran 10,080 reactions to improve a drug discovery coupling reaction (mean yield +52%, from 16.6% to 25.2%);[17]OpenAI — AI Chemist and LifeSciBench, 750 expert-authored tasks graded with 19,020 rubric criteria, where GPT-Rosalind passes 36.1% of tasks vs. GPT-5.5's 25.7%.[18]OpenAI — LifeSciBench
GPT-5.4 identified TEMPO as an additive improving Chan-Lam coupling of primary sulfonamides with boronic acids — a reaction important for drugs in oncology and infectious disease that historically gave low yields. The system generated proposals, designed experiments, analyzed data, and proposed follow-ups over 3 months (March 4 → June 4). Key metrics: 88% of boronic acids improved, 83% of sulfonamides improved; reactions exceeding 30% yield tripled (15.6% → 37.5%). Human chemists validated at bench scale: 11 of 14 substrate pairs confirmed, with 2x+ yield increase for most. Described as "near-autonomous" — humans still steered and corrected.
750 tasks across 7 workflows (evidence handling, analysis, design/optimization, scientific reasoning, validation/operations, translation, communication) and 7 biological domains. 173 PhD scientists authored the tasks; 453 independent expert reviewers validated them. 79% of tasks require multiple reasoning steps; 53% require interpreting artifacts (figures, PDFs, sequence files). GPT-Rosalind's strongest gains: Scientific Communication (56.3% → 71.1%) and Translation/bench-to-bedside (36.8% → 57.7%). Weakest: Design/Optimization (30.7%) and artifact-heavy tasks (28.1% pass rate).
A Nature study found Google's AMIE (Articulate Medical Intelligence Explorer) matched 21 primary care physicians in overall management reasoning and outperformed them in plan preciseness and guideline adherence, in blinded evaluations using patient actors in realistic long-term disease management scenarios.[19]Google Research
AMIE was tested specifically on long-term disease management — using drug formularies and clinical guidelines — rather than just diagnosis. Medical specialists evaluated both AMIE and the physicians using standardized scenarios with patient actors. Google's next steps: real-world clinical feasibility testing and a nationwide virtual care study. The implication is AI supplementing physician work on management tasks, freeing doctors for direct patient interaction.
Nate B Jones argues the WWDC question "did Siri get smarter?" misses the point. Siri is just the face; the product is an OS that can "take the hint" — screen context, personal files, app actions, Spotlight semantic index, Apple foundation models, and private cloud compute as an integrated agentic stack that doesn't spray your life to a random cloud service.[20]Nate B Jones
Apple's strategy: make the operating system itself feel agentic — can it find the file, move it, watch the page, build the shortcut, draft the message in the app where it'll actually be sent? Apple's technical stack: personal context + screen awareness + app actions + Spotlight semantic index + Apple foundation models + private cloud compute. This is what ordinary consumers want — "my aunt doesn't want to manage six models." The bet is that privacy-preserving on-device agentics is a stronger consumer proposition than the best chatbot.
Ramp argues organizations simultaneously overspend on AI (using expensive frontier models for routine tasks) and underuse it (avoiding frontier models for genuinely hard problems). Token-based budgeting hides this: $100K buys 5B tokens of advanced models or 210B of open-weight alternatives — volume tells you nothing about value.[21]Ramp Builders
1. Measure tasks, not tokens. Track atomic work units (invoices coded, tickets resolved, PRs reviewed) to calculate ROI. 2. Implement smart defaults. Route routine work to cheaper models that meet quality thresholds; set reasoning to medium, escalate only when necessary; use batch/flex modes for non-time-sensitive work (50% cost savings available). 3. Spend strategically on frontier models only for novel, ambiguous, high-stakes problems where additional intelligence meaningfully changes outcomes.
A big AI bill does not mean you are using too much AI. It means you are buying it wrong.
Simon Willison highlights Charity Majors' counterintuitive argument: because AI made code generation effectively free and instant, the discipline around what to build, how to review it, and how to maintain it becomes more important, not less. When lines of code are disposable and infinitely regeneratable, engineering judgment is the scarce resource.[22]Simon Willison
"The economics of code production were turned upside down. Instead of being very hard, time-consuming, and expensive to generate code, it became effectively free and instant."
Majors' full argument (from her article "AI demands more engineering discipline. Not less"): where developers once carefully curated and reused code, lines became disposable and easily regeneratable, fundamentally altering how engineering teams should approach their craft. The implication for teams: invest more, not less, in architecture decisions, code review rigor, and long-term maintainability.
YC partner John outlines a framework: don't overthink (both "searching for the perfect idea" and "self-disqualifying on founder-market fit" are failure modes); commit fully to one idea (burn the other boats — GovDash pivoted 5 times with full rebrands each time); and for the AI era, the best ideas sit at the edge of what models can do today, verticalize toward owning outcomes not selling software, and are "the most ambitious version of themselves."[23]Y Combinator
The bar for going deep on a customer segment isn't number of conversations — it's whether you could step into the customer's role and run their business tomorrow. Example: Corgi Insurance acquiring a carrier during their YC batch to own the outcome, not just sell software.
Even failed ideas generate unambiguous customer data and often surface deeper structural opportunities. The closing metaphor: navigating "early idea fog" requires moving forward even with limited visibility.
Google DeepMind's Logan Kilpatrick (in a Sequoia clip) describes how AI has changed his relationship to projects: previously, ideas felt "slightly out of reach" and he'd move on; now the opposite — ideas feel like they need to be even more ambitious because the technology enables them.[24]Sequoia Capital Lenny's podcast features a counterpoint: serial founder Mark Pincus argues the starting point should be "embarrassingly small" — Zynga succeeded because he was willing to make a Facebook poker app when he'd already run two companies.[25]Lenny's Podcast
"I feel like I used to kick around ideas and they were like slightly out of reach and I would just be like, 'Ah, wouldn't it be nice?' And now I have the opposite problem, which is I'm kicking around an idea and I'm like, 'I could probably make this even more ambitious.'" — Logan Kilpatrick
Pincus's complementary (not contradictory) point: once you have product-market fit, go big; but before PMF, multi-time founders fail precisely because they raise too much and recruit teams around a big vision before earning the right to scale. Lower your altitude to 1,000 feet, not 100,000 feet.
Fireship traced a causal chain through 10 landmark CS papers — Turing (1936), Shannon (1948), Rosenblatt (1957), Minsky/Papert (1969), Lamport (1978), Hinton (1986), Page/Brin (1998), AlexNet (2012), Attention Is All You Need (2017), GPT-3 (2020) — landing on a hot take: ChatGPT is structurally doing exactly what Shannon described in 1948 (predicting the next token), and Google's biggest mistake was publishing the transformer architecture openly.[26]Fireship
Key structural moments: Minsky/Papert's 1969 critique of single-layer perceptrons proved they couldn't learn XOR — but buried in the fine print was the acknowledgment that stacking layers would fix it. Lamport's 1978 logical clocks are now the bedrock of databases, blockchains, and multi-GPU training runs. Google assembled the largest human-text corpus ever built (via PageRank/web crawl), which accidentally became AI training data. AlexNet (2012) dropped ImageNet error rates by 10 points in one year, proving deep learning at scale. None of these inventors set out to build AI.
June 17's GitHub Trending includes GateGPT — a full transformer in Verilog running on a $50 FPGA at 80 MHz producing 56,000 tokens/second — plus a cluster of Claude Fable-adjacent tools: Fable Mode, Fable Codex, Fableize, and Fusion Fable. Omnigent (an open-source meta-harness for running Claude Code and Codex in parallel with budget caps) also trended.[27]Github Awesome
Other notable trending repos: xalen (data science), headroom (LLM memory optimization), gajae-code (AI agent tooling), guard-skills (agent safety guardrails), OpenHack (security), Polypost (content/social cross-posting), KU Kubernetes TUI (terminal dashboard), Caliper (ML benchmarking), and Inferoa (LLM inference optimization).[27]Github Awesome
Snap launched Specs — standalone AR glasses with a full AR display equivalent to a 24-inch monitor, dual Snapdragon chips, ~4 hours battery life — at $2,195. Snap stock dropped 9.6% on announcement day, with critics pointing to the price and bulky design vs. Meta's Ray-Ban Display at $799.[28]Tech Brew CEO Evan Spiegel called it "the computer for your face."
Specs capabilities: video streaming, gaming, navigation, and real-time environmental information. Pre-orders open in US, UK, France; shipping expected fall 2026. The real challenge isn't technical — it's whether consumers will wear conspicuous AR hardware at this price. Meta's Ray-Ban Display launches at $799 with a dramatically lower footprint.
"Almost 20 years since the launch of the iPhone, people are ready to think about computing differently." — Evan Spiegel
Lionsgate shares surged 14% on Netflix acquisition rumors (the stock is up 77% since January, owning John Wick and Hunger Games). Fox is acquiring Roku for ~$22B in a deal Netflix reportedly considered then withdrew from over antitrust concerns. IMAX hit record highs driven by Christopher Nolan's "The Odyssey" pre-sales; theater stocks broadly gaining as box office nears 2019 levels.[29]Sherwood News (Snacks)
McDonald's revived the fried apple pie after 30+ years (the baked version it replaced in 1992 has 10 more calories than the original). SpaceX's market cap passed Amazon's. Meta's Threads hit 500M monthly active users. Apple plans camera-equipped AirPods next year.
OpenRouter published a tutorial connecting Kilo Code (a VS Code/terminal coding agent with no built-in models) to OpenRouter's 300+ model catalog via a single API key. Configuration lives in kilo.json shared between extension and CLI; provider routing is customizable with sort, order, and only fields. Free tier: 50 req/day, scaling to 1,000 with $10+ in credits.[30]OpenRouter
A Nerd Snipe clip captures the economics of running frontier models in parallel agent workflows: while a single Mythos 5 session on the $200 plan is manageable, workflows spin up 8+ parallel agents simultaneously. Community examples: one user accidentally spun up 72 instances at once; another ran 250 Fable 5 instances split across two accounts to complete a single workflow.[31]Nerd Snipe
"I told it to use Ultra Code, and I wasn't paying too much attention. Apparently, it had spun up 72 at once."
"I had a workflow that did 250 [Fable instances]."
This is an under-discussed cost and rate-limit challenge as multi-agent workflows become standard. The account-switching workaround Theo demonstrated (in the Claude Code topic above) addresses part of this: since each tool call is a fresh API request, switching accounts mid-workflow routes future usage to the new account.
Robert Erez (~employee #8-9 at Octopus Deploy) joins Gergely Orosz for a deep dive on CI/CD reality vs. conference dogma: GitOps's four pillars never require Git (and secrets should never live in Git); "always roll forward, never roll back" is the right mental model once schema/state is involved; and feature toggles beat canary/blue-green for most app delivery scenarios.[32]The Pragmatic Engineer
~00:00 Introductions — Erez's background, ex-Skype-for-Web with host, now Octopus Deploy SaaS platform.
~10:00 The CI → CD → Continuous Deployment ladder: most teams only need to go up one rung.
~20:00 GitOps critique: four pillars (declarative, versioned/immutable, pull-not-push, continuously reconciled) never require Git specifically. Secrets in Git is an antipattern.
~40:00 Rollbacks: "Always roll forward" — rollbacks are a trap once schema/state is involved; teams claiming easy rollbacks are lucky.
~50:00 Feature toggles: granular, precise, instant — but breed stale-flag hygiene problems if not managed.
~60:00 Kubernetes on-prem including clusters on research vessels at sea.
~70:00 Platform teams and the Octopus "reef" — VMs-to-Kubernetes SaaS rebuild; ~200 days for 50% adoption.
~80:00 AI's impact on CI/CD: more code velocity expected; pipeline speed de-emphasized (agents babysit builds); focus shifts to risk reduction via progressive delivery and agent-flippable feature toggles.
"The problem with GitOps is everybody's talking about it but there's no single agreed-upon definition."
Joseph Krause (CEO, Radical AI) argues AI can generate novel materials compositions well, but cannot one-shot a real material — synthesis, characterization, processing, manufacturing, and 10-year qualification pipelines are not solvable by models alone. Inorganic materials can't be string-encoded like molecules in bio. There is no AlphaFold for materials science. Self-driving labs solve this by running entire research campaigns autonomously — not just high-throughput experiments.[33]Latent Space
~00:00 Who is Joseph Krause — materials scientist, CEO of Radical AI (~2.5 years old), focused on structural metals and high-entropy alloys.
~10:00 Why materials science ≠ drug discovery: no SMILES/SELFIES string encoding, supply chain and cost and microstructure can't fit in a string.
~20:00 Self-driving lab vs. automated lab — the Waymo analogy: a self-driving lab runs entire research campaigns, not just high-throughput experiments.
~30:00 Three pillars of Radical AI's platform: hard physical challenges, OS-layer software, and automation/connection.
~40:00 Experiments as the moat, not models — why Radical AI open-sources Matrix and Torch Sim.
~50:00 Hot take: tool vendors should rebuild tools for agents/robots. ML engineers should stay MLEs working in science, not become materials scientists.
~60:00 On China: copy the focus not the single-entity ownership — win via workforce + public-private partnership.
"We are experiment-constrained, not compute-constrained. The experiment is the moat."
The full-week AI recap covers: Fable 5/Mythos standoff with the US government entering its second week; Apple's Siri AI strategy at WWDC (on-device agentic OS vs. chatbot race); SpaceX IPO and Cursor acquisition; OpenAI's leaked financials ($38.5B net loss mostly accounting, $73B cash, strong inference margins); and a biosecurity open letter warning frontier models can substitute for expert knowledge in bioweapon development.[34]Last Week in AI #248
~00:00 Fable 5 release recap and access suspension timeline.
~10:00 Siri AI and WWDC: on-device intelligence vs. cloud chatbot.
~20:00 SpaceX IPO surge, Cursor acquisition, Elon becomes first trillionaire.
~30:00 AI IPO wave: Bezos's Prometheus $12B raise, DeepSeek $7B raise.
~40:00 AI policy: government equity stake proposals, Dario's policy essay, Anthropic's RSI/global-pause post.
~50:00 Biosecurity open letter: frontier models can substitute for "novel bioweapon" expert knowledge. "Fills me with dread."
~60:00 Open-source releases: Gemma 4 12B, Diffusion Gemma (4x faster text generation), Huawei post-training DeepSeek's 1.6T model on Ascend chips.
~70:00 Musicians vs. labels: AFM lawsuit against UMG/WMG over AI training data.