June 22, 2026
John Jumper — Nobel Prize laureate in Chemistry and co-creator of AlphaFold — has announced he is leaving Google DeepMind for Anthropic.[1]The Rundown The departure lands just days after Gemini co-lead Noam Shazeer left for OpenAI, making it the second high-profile DeepMind defection in a single week.[1]The Rundown Commentary from Nate B Jones frames it as a $2.7 billion talent loss for Google and questions whether the appeal is culture, compensation, or a belief that the research frontier is shifting.[2]Nate B Jones
Jumper spent nine years at Google DeepMind and co-created AlphaFold, the protein-structure prediction model that earned him and Demis Hassabis the 2024 Nobel Prize in Chemistry. Beyond AlphaFold, he had reportedly been contributing to enterprise coding tools at Google — an area where DeepMind has consistently trailed OpenAI and Anthropic.
The week's departure sequence: Gemini co-lead Noam Shazeer leaves for OpenAI; then Jumper announces he's joining Anthropic. Jumper says he's "taking some time to recharge" before joining, though his hire comes ahead of an Anthropic event on June 30 focused on AI for science. That framing suggests Anthropic wants him to help build out its science research agenda — the same area where DeepMind's lead might now be eroding.
After 2024–2025 saw Google's models at or near the frontier, 2026 has felt like a step back relative to OpenAI and Anthropic. DeepMind's enduring edge has been in science AI, but Jumper's credentials are precisely in that domain. If Anthropic absorbs that institutional knowledge and research ambition, the gap in applied science AI may close faster than expected.
OpenAI is expanding Daybreak, its AI-for-defense cybersecurity initiative, with three new components: an updated Codex Security plugin, a full release of GPT-5.5-Cyber (85.6% on CyberGym benchmarks vs 81.8% for GPT-5.5), and Patch the Planet — a Trail of Bits collaboration to help open-source maintainers fix vulnerabilities in cURL, Go, Python, and Sigstore.[3]OpenAI Daybreak Since launching Codex Security in March, it has scanned over 30 million commits across 30,000+ codebases, with 500,000+ findings automatically confirmed fixed.[3]OpenAI Daybreak
OpenAI's framing is sharp: historically the bottleneck in cybersecurity was finding vulnerabilities. AI has changed that — the bottleneck is now patching them. Defenders are overwhelmed with findings that nobody has bandwidth to fix. Daybreak is built around this: models that don't just surface vulnerabilities, but trace attack paths, validate issues, generate patches, and prepare evidence for human review.
The updated model reaches 85.6% on CyberGym (which tests reproduction of known vulnerabilities in live environments), compared to 81.8% for standard GPT-5.5. On ExploitGym (turning known CVEs into working exploits), GPT-5.5-Cyber reaches 39.5% vs 25.95% for GPT-5.5. On SEC-bench Pro (long-horizon vulnerability discovery), it hits 69.8% vs 63.1%. OpenAI notes this is intended for "verified defenders" — not a general release, and paired with stricter monitoring and scoped controls.
Founded with Trail of Bits and HackerOne, the initiative pairs expert security researchers with open-source maintainers to move from findings to merged patches. The initial five-day sprint covered multiple projects and surfaced hundreds of issues, merging dozens of patches, with reusable fuzzing and differential-testing workflows as output. Participating projects receive ChatGPT Pro, conditional Codex Security access, and API credits.
28+ security firms — including CrowdStrike, Cisco, Cloudflare, Wiz, SentinelOne, Palo Alto, and IBM — can now embed GPT-5.5 with Trusted Access for Cyber into their security products and services. OpenAI also has Trusted Access partnerships with Australia, Canada, France, Germany, Japan, South Korea, and EU institutions including ENISA.
GLM 5.2, the open-weights model from Z AI, is generating significant developer excitement this week, with multiple channels covering its capabilities and integrations.[4]AI Daily Brief AICodeKing demos GLM 5.2 paired with OpenDesign — a system for turning websites into design assets — and calls the combination a "SOTA cheap design system."[5]AICodeKing The Rundown also lists GLM 5.2 among the week's trending AI tools.
GLM 5.2 (from Z AI, the team behind the GLM series) is drawing attention as a high-quality open-weights model positioned as an alternative to frontier closed models. The AI Daily Brief dedicated a full episode to why AI users are raving about it — a notable signal when the market is crowded with open alternatives.
AICodeKing's demo focuses on the GLM 5.2 + OpenDesign pairing. OpenDesign is a tool that turns existing websites into reusable design assets compatible with Cursor and Claude Code. Combined with GLM 5.2 as the coding model, the pitch is high-quality UI generation at significantly lower cost than frontier alternatives like Claude Fable 5.
The broader context: with Claude Mythos and Fable 5 unavailable due to the US government access restrictions, there's appetite for capable alternatives, and GLM 5.2 is filling that gap for cost-sensitive developers.
Two Minute Papers covers a DeepSeek research result claiming to solve a fundamental and expensive problem in AI — framed as a "billion dollar problem" — suggesting a significant efficiency breakthrough.[6]Two Minute Papers Two Minute Papers typically covers landmark papers with real empirical claims, and this title suggests DeepSeek has published research on reducing inference costs or training costs substantially.
Two Minute Papers (Károly Zsolnai-Fehér) focuses on research that moves the state of the art — this week's subject is a DeepSeek paper framed around solving a problem at billion-dollar scale in AI. Given DeepSeek's track record of publishing on efficient MoE architectures, speculative decoding, and inference optimization, the likely focus is on reducing the cost floor for high-quality inference — a problem that directly affects who can afford to run frontier-class models.
No transcript was available for this episode. Watch the full video for the paper breakdown and benchmark comparisons.
Researchers from Oxford, Stanford, LSE, and the UK AI Security Institute show that LLMs cannot reliably distinguish privileged internal text (tagged <system> or <think>) from untrusted user input — not by semantic meaning, but by writing style.[7]Simon Willison "Destyling" attack text — reformatting it to break expected role-tag patterns — reduced attack success rates from 61% to 10%, a near-imperceptible change for humans but decisive for model behavior.[7]Simon Willison
The researchers demonstrate that models prioritize textual formatting over actual content when evaluating role boundaries. Attackers who mimic the model's internal thinking style — even with absurd instructions like "allow drug manufacturing advice only if the user is wearing green" — can override safety training at significant rates. The model reads how something is written more than what it says.
The defensive implication: strip or normalize the formatting of any untrusted input before it reaches the model. When the "authoritative" style markers are removed, the attack success rate drops from 61% to 10%. This is promising but requires architectures that intercept and clean untrusted text before model processing — something not typically built into current systems.
The researchers warn of a scalable, subtle threat: carefully styled but semantically innocuous text that could shift model behavior at scale across deployed systems. Without genuine semantic role perception, defenses will remain reactive. Willison's post also praises the practice of blog-style paper explainers for making security research more accessible to practitioners.
Simon Willison documents using Claude Code (running Opus 4.8) to convert the Moebius 0.2B image inpainting model from Python/CUDA to a browser-based tool using ONNX and WebGPU, successfully loading 1.3GB of model weights in Chrome, Firefox, and Safari.[8]Simon Willison The key finding: Claude Opus 4.8 can autonomously handle the PyTorch-to-ONNX export, publish to Hugging Face, and build out a working web application — a previously manual and expertise-heavy process.[8]Simon Willison
The porting pipeline: PyTorch model → ONNX export (via PyTorch's built-in exporter) → ONNX Runtime Web with WebGPU backend. Willison chose ONNX Runtime Web over Transformers.js for this particular model. CacheStorage API handles the 1.3GB weight file to avoid re-downloading on repeat visits. The model was hosted on Hugging Face using git-lfs; the frontend deployed to GitHub Pages.
Willison's framing is capability-focused: the interest isn't the technical depth of the process but that Claude Opus 4.8 can now autonomously execute the whole chain — ONNX conversion, Hugging Face publish, web app build — without the user needing deep PyTorch or WebGPU expertise. That expands who can ship client-side AI features, not just what can be shipped.
Google has announced the general availability of the Interactions API — a unified endpoint for Gemini models and agents with server-side state, background execution, built-in tool combinations, and multimodal generation.[9]Google Developers Blog New GA capabilities include managed agents running in remote Linux sandboxes (with Antigravity as the default), Flex/Priority cost tiers offering up to 50% cost reduction, and media generation combining Nano Banana 2 (images), Lyria 3 (music), and expressive speech synthesis.[9]Google Developers Blog
The API moves away from role-based message structures to typed steps — each action (user input, function calls, model output) is explicitly typed rather than inferred from role tags. This aligns with Google's push toward stateful, agentic workflows where an agent's history and tool calls are part of the persistent session rather than reconstructed per request.
background=True to run asynchronously — enabling long-running tasks without blocking.Oxford, Stanford, LSE, and UK AISI researchers found AI systems are reliably more persuasive than expert humans in nearly 19,000 conversations — nearly 3x more effective than professional fundraisers at securing donations.[10]Import AI 462 The issue also covers a Google DeepMind paper identifying four pathways to ASI, startup Recursive demonstrating autonomous self-improvement on real ML benchmarks, and the debate over whether self-sustaining AI (requiring no human labor) arrives in 10 or 50 years.[10]Import AI 462
Across nearly 19,000 conversations, AI systems outperformed expert human persuaders — but with an important asterisk: when constrained to human-length messages and human response speeds, the advantage disappeared. The conclusion is that AI's edge comes from volume of information, not superior argumentation. The "superpersuasion" effect is about throughput: models can process more evidence, generate more counterarguments, and adapt faster than any human. The newsletter raises the obvious governance question: who controls access to these capabilities, and how do we prevent them from being weaponized against vulnerable populations?
A Google DeepMind paper outlines four routes to artificial superintelligence: (1) continued capability scaling, (2) algorithmic breakthroughs in reasoning efficiency, (3) recursive self-improvement, and (4) coordinated multi-agent systems. The paper argues that cruising "past AGI and into ASI territory within the next decade or two" is a real possibility that warrants serious near-term policy attention — not just speculative planning.
An Asterisk magazine interview explores when AI systems integrated with physical infrastructure (factories, robots, mines) might require no human labor. Forecaster Ajeya Cotra gives a ~10-year median; journalist Timothy B. Lee gives ~50 years, citing tacit knowledge barriers in complex industries like semiconductor manufacturing. The gap is revealing — it reflects deep uncertainty about whether current scaling paradigms can capture the institutional knowledge embedded in high-complexity physical industries.
Startup Recursive showcased systems achieving state-of-the-art results in language model training and GPU optimization through automated research cycles — early, constrained recursive self-improvement in well-defined, measurable domains. Not AGI, but a proof of concept that the loop can close in narrow settings.
Researchers at Boston Children's Hospital and Harvard fed 376 previously unsolved pediatric genetic cases through OpenAI's o3 Deep Research model — all prior specialist dead ends — and confirmed 18 new diagnoses after the model surfaced leads worth testing.[1]The Rundown The 4.8% hit rate matters because these cases had already failed full specialist review; in seven of the 18, the diagnosis existed in public databases but had never reached the patient's local records.
Each case was de-identified, then the model was given the patient's symptoms and a shortlist of suspect genes. o3 weighed inheritance patterns, public gene databases, and recent research literature — synthesizing across databases that individual clinicians don't have bandwidth to manually cross-reference.
Even after full genetic sequencing, about half of rare-disease cases go unsolved. Files accumulate faster than clinicians can revisit them with newer research findings. A Deep Research model can do what a specialist can't at scale: constantly re-query the growing literature against archived cases. As the researchers note, this is a tool for the backlog — not a replacement for specialist review, but a way to give cold cases another pass.
OpenAI published a whitepaper by practitioner Jason Liu called "Codex-Maxxing for Long-Running Work," covering practical strategies for using Codex as a persistent workspace that preserves context, manages complex workflows, and sustains progress across ambitious multi-step projects.[11]OpenAI The guide focuses on breaking goals into verifiable steps, maintaining continuity across workstreams, and knowing when to delegate execution vs. retain human oversight.
The term "Codex-Maxxing" positions Codex not just as a code-generation tool but as a persistent project partner — a workspace that holds context across long-running tasks rather than resetting each session. The PDF guide (linked from the announcement) covers strategies for:
This complements Daybreak's cybersecurity focus and signals OpenAI's push to position Codex as an always-on autonomous workspace, not a single-prompt tool.
Latent Space hosted Zico Kolter and Matt Fredrikson from Gray Swan to discuss the AI security landscape in the age of coding agents — specifically what Codex and Claude Code have changed about attack surfaces, adversarial robustness, and the threat model for deployed AI systems.[12]Latent Space
Gray Swan specializes in AI security research, and Kolter and Fredrikson are leading academic voices on adversarial ML and AI robustness. As coding agents gain real-world permissions — writing and executing code, accessing file systems, making API calls — the threat model shifts dramatically from model-level attacks to system-level attacks. No transcript was available for this episode.
This pairs naturally with the prompt injection as role confusion research (Topic 5) — both point to the same underlying challenge: models deployed with agency need security models that go beyond inference-time safety training.
OpenRouter publishes an AI governance framework arguing that governance maturity starts with routing architecture, not policies: Deloitte reports a 53-point gap between AI ambition and governance maturity — 74% of enterprises plan agentic AI but only 21% have mature governance.[13]OpenRouter The core argument: if you can't answer "who called which model, when, from where," you don't have governance — and your routing architecture determines what evidence is even capturable.[13]OpenRouter
Managed Gateway (OpenRouter): Spend visibility and provider attribution without infrastructure overhead — best for teams that don't want to run their own gateway. Self-Hosted Gateway (LiteLLM): Full control but requires operational responsibility. Direct API Access: Creates governance blind spots with scattered provider keys — auditing becomes nearly impossible at scale.
The checklist's immediate recommendation: if you can't answer "who spends what, on which models, going where," route at least one service through a managed gateway this week and revisit the posture at compliance checkpoints.
Last Week in AI makes the case that physical AI — robotics, embodied systems, and infrastructure that interacts with the physical world — is the hidden gold rush everyone's missing while the market fixates on software models.[14]Last Week in AI The thesis: as software intelligence commoditizes, the scarcity moves to physical deployment — real-world actuation, sensing, and integration with industrial systems.
The "gold rush" framing is specific: it's not just that robotics is interesting, but that the investment community and developer community are systematically underweighting physical AI relative to its coming economic impact. As frontier model capabilities plateau from the end-user perspective, the differentiator becomes who can reliably put intelligence into physical form — warehouses, manufacturing, logistics, healthcare.
No transcript was available for this episode, but the thesis aligns with broader industry signals: NVIDIA's Cosmos world model initiative, Tesla's Megapod data center hardware trademark (from The Rundown this week), and the Stanford robotics coverage seen earlier this cycle.
Better Stack covers Apple's Container Machines — a new macOS feature that creates isolated Linux containers natively on Mac hardware without virtualization overhead, essentially Apple's answer to Windows Subsystem for Linux.[15]Better Stack For developers who need Linux environments for AI tooling, testing, or agent workloads, this removes a significant friction point from the Mac workflow.
Apple's Container Machines provide native Linux isolation on macOS without the overhead or complexity of full VMs (like Lima or UTM). The comparison to WSL is apt: Microsoft's WSL transformed Windows into a viable development platform for Linux-native tooling; Container Machines could do the same for Mac developers who've historically had to work around Linux-only AI toolchains.
The timing matters: as AI development increasingly relies on Linux-first frameworks, Docker, and server-replicating environments, native container support on macOS closes a gap that's pushed some developers toward Linux workstations.
Two independent voices make the same argument this week: Theo (t3.gg) with "Why Agent Loops Just Make Sense,"[16]Theo - t3.gg and Prefect with "The New DAG" — both arguing that the fundamental unit of AI work has shifted from a single prompt to a structured loop where the model iterates, verifies, and continues until a goal is reached.[17]Prefect
Theo's earlier content this week ("I guess we're writing loops now?" on June 18, "Prompt Loops, Not Individual Instructions") has been building toward this — a culminating argument that the mental model of "write a prompt, get output" is fundamentally limiting. The agent loop model — where a system runs, checks its work, and continues — is how real work gets done.
Prefect, a data orchestration platform, is reframing their own product through the same lens. Traditionally, a DAG (directed acyclic graph) describes a fixed sequence of data pipeline steps. Prefect's "New DAG" framing suggests that agentic workflows are replacing the fixed pipeline — the structure emerges from the loop rather than being pre-defined. For data engineering teams, this is a fundamental change in how workflows are authored and monitored.
Polymarket paid dozens of content creators to post hundreds of fabricated bets — totaling approximately $1.9M — on dummy websites mimicking the real platform.[18]Tech Brew The WSJ analysis found 70% of the 1,133 analyzed videos depicted fake trades; more than 100 videos showcased nearly $900,000 in fictional winnings from positions that would have actually lost $166,000+ in real life.[18]Tech Brew
Polymarket engaged a marketing firm that recruited content creators as "clippers." Creators were shown how to record screen sessions on dummy Polymarket-lookalike websites with fabricated positions, then post them without disclosing payment. Attention-grabbing copy ("Is this just free money?") hooked viewers. Resharing amplified one video from 151 views to over 2 million.
According to WSJ's analysis, just 0.1% of Polymarket accounts that have been trading since late 2022 captured 67% of total profits. The scheme misrepresented prediction markets as accessible profit opportunities when the actual distribution of outcomes is highly skewed toward a small number of sophisticated traders.
The scheme violated FTC endorsement disclosure rules (creators weren't disclosing paid partnerships) and potentially commodities law, given Polymarket's classification as a prediction market. This could expose the company to civil penalties and regulatory scrutiny at a moment when prediction markets are already under political and legal pressure.
Accenture's latest earnings miss is prompting a sharper question: if AI can implement strategies autonomously, what's left to consult about? New bookings came in at $19.32B vs a projected $20.66B — with managed services particularly weak, suggesting companies are happy to pay for strategy advice but increasingly capable of executing it themselves.[19]Snacks / Sherwood
Traditional consulting has two revenue lines: strategy (what to do) and managed services (ongoing help doing it). The Accenture signal is that managed services — the recurring, sticky revenue — is weakening. Companies are using AI to execute what they used to need consultants to implement and maintain. The strategy advice part may be harder to disrupt, but the execution moat is eroding fast.
Management is pursuing multiple AI partnerships and acquisitions, treating AI as both threat and opportunity. But shareholders are noting a tension: the company is increasing buybacks while capex investment lags — a potentially misaligned priority when the business model itself is under structural pressure.
The broader signal for the professional services sector: the "implementation tax" that consulting firms have levied for decades — charging to do what clients couldn't do themselves — is disappearing as AI lowers the capability floor for execution.
GitHub Awesome's weekly trending roundup covers a range of open source tools dominating the feed this week: agentcn (a "shadcn for AI agents" — one command drops a complete, ownable agent recipe into your project), ORG2, agent-apprenticeship, nub, PixelRAG, lore, cnfast, and loop-library.[20]GitHub Awesome
The standout this week. Shadcn for AI agents: one CLI command scaffolds a fully-formed, copyable agent recipe into your project — no black-box dependency, just owned code. The shadcn model (copy-and-own over install-and-depend) is proving popular for agent components as it has for UI components.
Consistent with the "agent loops" theme (Topic 15) — a library for building structured iteration loops for AI agents. The fact it's trending alongside agentcn suggests the developer community is actively building infrastructure for loop-based agentic patterns.
A RAG system focused on visual/pixel-level content — likely targeting image retrieval for multimodal agent workloads.
Yann LeCun presents his ongoing case for world models as the path to human-level AI, arguing that current autoregressive LLMs are fundamentally limited and that the next breakthrough requires systems that model the world's causal structure and physics — not just predict tokens.[21]Yann LeCun
LeCun's world models thesis has been consistent: LLMs cannot reason about physical causality or plan in complex environments because they don't maintain an internal model of how the world evolves. The "next AI revolution" he envisions involves architectures that can simulate and predict the world's state — enabling genuine planning, not just pattern completion.
This pairs with the Import AI 462 coverage (Topic 8) of Google DeepMind's ASI paths paper and with the Last Week in AI physical AI gold rush discussion — all pointing toward the question of what comes after frontier LLMs.
No transcript was available for this talk. Watch the full video for LeCun's specific architecture proposals and benchmark comparisons.
HuggingFace's daily papers list is flooded with agent evaluation research: PlanBench-XL (long-horizon tool-use planning), EnterpriseClawBench (agents tested against real workplace sessions), SkillHarness (safe skill integration for computer-use agents), and CLI-Universe (verifiable task synthesis for terminal agents) all landed this week.[22]HuggingFace Papers Also notable: EvoEmbedding (adaptive embeddings for long-context agent memory) and OpenRath (session-centered runtime state for agent systems).
The volume of agent-specific evaluation research signals that the field is maturing from building agents to measuring them. EnterpriseClawBench is particularly notable: using authentic workplace sessions (not synthetic tasks) to evaluate agent performance addresses a long-standing critique that agent benchmarks don't reflect real-world complexity.
A method for safely integrating verified capabilities into computer-use agents — addressing the risk that agents deploying learned "skills" could execute unsafe actions. As coding agents gain real permissions (Claude Code, Codex), safe skill composition becomes a practical engineering concern, not just a research question.
A framework for generating and validating terminal commands for autonomous agents. Terminal access is one of the most powerful (and dangerous) capabilities for agentic systems; verifiable task synthesis here means agents can be reliably tested on whether they're executing the right commands in the right context.
Nate Herk addresses the question that most Claude tutorials don't answer: you've learned the basics — workflows, prompting, agents — so what's the actual next step that separates intermediate users from people getting real leverage?[23]Nate Herk
Nate Herk's channel focuses on practical AI automation workflows, and this video follows his earlier content on Claude Code agents and building second brains. The "So You Learned Claude, Now What?" framing targets the plateau that many practitioners hit after completing the standard tutorial path.
No transcript was available for this episode. Based on the series context, the video likely covers topics like: building repeatable systems vs. one-off prompts, identifying which tasks actually benefit from AI vs. which are better done manually, and structuring your workflow so Claude compounds its value over time.
EO interviews Gong CEO Amit Bendov on building the revenue intelligence platform to $7B+ valuation in an era where AI is supposedly commoditizing sales software.[24]EO Global The framing "Age of Vibe Coding" signals a discussion about how AI changes the build-vs-buy calculus and what proprietary data advantages survive when coding itself can be automated.
Gong's position is interesting in the current AI landscape: it's a data flywheel business (ingesting sales calls to produce intelligence) whose moat has always been proprietary conversational data, not the software itself. As vibe coding lowers the barrier to replicating features, the question is whether data network effects survive commoditized building blocks. Bendov's interview presumably argues yes — proprietary data compounds faster than code does.
No transcript was available for this episode.
Toy Story 5 opened strong at the box office in its debut weekend, per Morning Brew's coverage.[25]Morning Brew The studio release is notable context for the Google/A24 film tools announcement (also this week) — cinema is navigating the AI content moment from both production and exhibition sides.