The Bleeding Edge

// Episode W22 · 2026-05-22 to 2026-05-29

Anthropic had the single most concentrated week of momentum any frontier lab has put together since GPT-4 launched

Anthropic had the single most concentrated week of momentum any frontier lab has put together since GPT-4 launched. Andrej Karpathy joined the company. Reporting says Anthropic is about to post its first-ever profitable quarter. Pope Leo XIV published the Vatican's first AI encyc…

The Bleeding Edge — Episode Briefing W22

Date range: 2026-05-22 to 2026-05-29 (Europe/Madrid)

Headline of the Week

Anthropic had the single most concentrated week of momentum any frontier lab has put together since GPT-4 launched. Andrej Karpathy joined the company. Reporting says Anthropic is about to post its first-ever profitable quarter. Pope Leo XIV published the Vatican's first AI encyclical, co-presented with Anthropic interpretability lead Christopher Olah. Claude Opus 4.8 — internally branded "Mythos" — began rolling out to all customers. In the same seven days: Google I/O 2026 dropped 100 product announcements, Sequoia told the market that the next trillion-dollar company won't sell software (it will sell the work), Meta cut 8,000 staff, and SpaceX's IPO filing exposed xAI's standalone finances as far worse than the headline valuation suggested. Capital, talent, regulatory cover, and now moral authority are flowing toward the same two or three labs — and the field's economic logic is visibly shifting from licensing software to delivering outcomes.

Top 5

  1. Andrej Karpathy joins Anthropic. The Creators' AI weekly digest broke that the OpenAI co-founder and ex-Tesla AI director — most recently running the Eureka Labs education project independently — is now at Anthropic. Why it matters: Karpathy is the highest-profile individual transfer in the industry's history; pairing arguably the field's most influential teacher with the lab that has been winning the interpretability and enterprise narratives compounds Anthropic's recruiting flywheel and weakens xAI and OpenAI's bench at the same time. Unverified Source: Creators' AI weekly digest.

  2. Anthropic is about to post its first-ever profitable quarter. The same digest reports Anthropic's next quarterly print will be the first in the black, driven by enterprise Claude seats and API consumption from agent platforms. Why it matters: the "frontier labs are bonfires of cash" framing has been the default investor assumption since 2023; a profitable Anthropic quarter — even on a non-GAAP basis — collapses that narrative and gives the company independent fundraising leverage versus OpenAI and xAI, both of whom remain deeply loss-making. Unverified Source: Creators' AI weekly digest.

  3. Sequoia: "the next $1T company won't sell software. It will sell the work." Sequoia published a position piece arguing the trillion-dollar company of the next decade will be an AI-native services firm that delivers business outcomes — not a SaaS vendor selling seats. In parallel, The AI Opportunity wrote up the Big Four's enterprise AI agent playbook, describing how Accenture, Deloitte, EY and KPMG have moved from one-off pilots to standardised agent factories selling against the SaaS stack. Why it matters: if Sequoia is right, the buyer category most exposed is horizontal SaaS; the buyer category most rewarded is consulting-adjacent agent delivery — which is where Sierra, Cognition, and the Big Four are all pointed. Inference Sources: The AI Opportunity — Sequoia writeup, The AI Opportunity — Consulting Firm Playbook.

  4. Google I/O 2026 drops 100 announcements. Google's annual developer conference shipped what the Creators' AI summary tallied at 100 product, model, and platform updates spanning Gemini, AI Mode in Search, Android XR, Workspace, Vertex, and developer tooling. Why it matters: the surface area is the story — Google is now executing on every horizontal AI surface simultaneously, which means buyers can no longer assume a "Google is behind" discount in procurement. Specific items will land throughout the rest of W22 and W23 coverage; the volume itself is what reframes the competitive picture. Unverified Source: Creators' AI weekly digest.

  5. Pope Leo XIV publishes the Vatican's first AI encyclical — co-presented with Anthropic's Christopher Olah. The Holy See released its first encyclical letter on artificial intelligence under Pope Leo XIV, with Anthropic's head of interpretability Christopher Olah named as a co-presenter at the formal launch. Why it matters: encyclicals are the highest-tier teaching documents in Catholic doctrine and shape the moral framing used by 1.4 billion Catholics — and a meaningful share of European policymakers. Having a frontier lab's interpretability lead alongside the Pope is an unprecedented legitimacy event for Anthropic specifically and for the "mechanistic understanding as safety" framing more generally. Unverified Source: Creators' AI weekly digest.

Categorised News

Frontier & Big Tech

Claude Opus 4.8 ("Mythos") rolling out to all customers. Anthropic confirmed Claude Opus 4.8 — referred to in product copy as the Mythos model — will be available to all customers "in the coming weeks." The release is the first major Opus refresh since 4.6 and slots in alongside the NLA interpretability work shipped in W19. Corroborated Sources: Anthropic, Capital Brief Standup.

Qwen 3.7, DeepMind Co-scientist, ByteDance Lance, Mega ASR, and LeRobot Humanoid all land in the same news cycle. AISearch's roundup catalogued five frontier launches in one week: Alibaba's Qwen 3.7, DeepMind's Co-scientist research agent, ByteDance's Lance multimodal model, a "Mega ASR" speech-recognition system, and a new LeRobot humanoid release. The volume itself is the data point — frontier launches are now a weekly cadence across at least four geographies. Unverified Source: AISearch on Substack.

Anthropic publishes the cleanest read yet on which jobs AI will replace first. Anthropic released a labour-market report breaking down task-level exposure to Claude across occupations, plus a breakdown of where "founder-shaped" opportunities are concentrated. The AI Opportunity summarised it as the most data-grounded read on AI labour impact since the 2023 OpenAI/Penn paper. Unverified Source: The AI Opportunity.

Market Cap / Valuation

SK Hynix becomes the first memory company to cross $1T market cap. Reuters reported SK Hynix's market capitalisation topped $1 trillion this week, driven by HBM demand from NVIDIA-class accelerators. Why it matters: memory is now the binding constraint on frontier training and serving, not logic — SK Hynix's print is the cleanest market signal of where the AI infra rent is actually accruing. Corroborated Source: Reuters via Capital Brief.

xAI's real finances revealed via SpaceX's IPO filing — and they're ugly. The Creators' AI digest flags that the SpaceX S-1 (filed this week ahead of the long-rumoured IPO) contains the first detailed disclosure of xAI's standalone financials, and that the numbers are materially worse than the headline valuation implies. Specific figures were not independently confirmed in the week's flow. Unverified Source: Creators' AI weekly digest.

Firmus closes a USD 505M equity raise — and NVIDIA's circular-financing model becomes visible. Australian AI infrastructure company Firmus raised USD 505M, with Capital Brief reporting NVIDIA participation alongside other strategics. The Brief frames the deal as the latest data point in NVIDIA's "circular financing" pattern — investing in customers who then commit to multi-year GPU purchases. Unverified Source: Capital Brief.

Meta cuts 8,000 staff. Meta announced a further 8,000 job cuts this week, the largest of its three rounds since 2023. The digest framed the cuts as concentrated in non-AI product lines, freeing capacity for the Reality Labs / AI superintelligence org. Unverified Source: Creators' AI weekly digest.

Apps / Dev Tools / Platforms

Claude Code Cowork goes mainstream. Lenny's How I AI ran a long interview with Anthropic's Felix Rieseberg on his actual workflow inside Claude Code Cowork — the multi-instance collaborative coding environment Anthropic has been rolling out. Rieseberg demonstrated using Cowork sessions for 3D house design and a USD 20 physical "AI buddy" hardware project, including the full prompting playbook. Unverified Source: Lenny's How I AI.

Personal-agent rebuild on Hermes + GPT-5.5 shows the agent stack is now portable. The Creators' AI newsletter published a detailed migration tutorial: rebuilding a personal agent that "runs two businesses in one Telegram chat" on Nous Research's MIT-licensed Hermes model paired with GPT-5.5 — moving off the prior OpenClaw setup. Why it matters: the practical signal is that agent personalities, memory, and tool integrations are now portable across model backends. The lock-in story for personal-agent platforms is weaker than it was six months ago. Unverified Source: Creators' AI on Hermes migration.

Infrastructure & Ecosystem

NVIDIA's Polar: token-faithful RL training for agent harnesses. NVIDIA open-sourced Polar, a framework that proxies LLM API calls during rollouts to reconstruct token-faithful trajectories — letting researchers run GRPO-style reinforcement learning against unmodified Codex, Claude Code, and Qwen Code harnesses. Using simple GRPO, Polar improves Qwen3.5-4B by 22.6 points on SWE-Bench Verified with the Codex harness. Why it matters: the bottleneck on agent RL has been that training-time trajectories don't match deployment-time tool-use; Polar collapses that gap and dramatically lowers the cost of producing strong open coding agents. Corroborated Sources: MarkTechPost, arXiv 2605.24220, GitHub NVIDIA-NeMo/ProRL-Agent-Server.

Together AI open-sources OSCAR — 2-bit KV cache, 8× memory reduction, near-BF16 accuracy. OSCAR is an attention-aware 2-bit KV cache quantisation system for long-context serving. Together claims roughly 8× KV memory reduction and up to 7× serving throughput at large batch sizes, while staying within near-BF16 accuracy on Qwen3-8B. Every prior INT2 approach hit accuracy cliffs on long-context tasks; OSCAR doesn't, per the paper. Why it matters: serving cost is the binding constraint on agent economics; this is the first credible 2-bit KV approach that actually preserves quality at production scale. Corroborated Sources: MarkTechPost, arXiv 2605.17757v1, GitHub FutureMLS-Lab/OSCAR.

EAGLE 3.1 fixes speculative-decoding attention drift. A new EAGLE iteration targets the attention-drift failure mode that broke deeper-step speculative decoding in earlier versions, restoring throughput gains on long-context generation. Bundles cleanly with OSCAR for end-to-end serving stack improvements. Corroborated Source: MarkTechPost.

Regions / Macro

US-China tentative trade arrangement signalled, then walked back. Treasury Secretary Scott Bessent declined to confirm a tentative US-China trade agreement at a press conference, even as multiple outlets reported one was close. The relevance for AI buyers: any movement on export controls, especially around HBM and advanced packaging, materially repricesthe NVIDIA / SK Hynix / TSMC stack. Unverified Source: Capital Brief Standup, press conference video.

AI in Consumer Hardware

LeRobot humanoid release lands quietly inside a busy week. Hugging Face's LeRobot project shipped its first humanoid form-factor release. Specs and pricing weren't surfaced in the digest, but the release continues the open-source robotics flywheel that has been building since W14. Unverified Source: AISearch on Substack.

Prompting Skill of the Week

Technique: The Cowork Critic Pattern. Best for: complex agentic work — code, research, writing — where you've noticed the model converges on a single answer too quickly and you can't tell whether it's right. Inspired by the multi-instance Claude Code Cowork workflows Felix Rieseberg detailed in this week's How I AI.

  1. In one window, give the model the task and ask for its best answer. Save the response.
  2. In a separate fresh window, paste the original task plus the first response as a "draft solution from a colleague," and ask the model to critique it ruthlessly, listing every weakness, edge case missed, and stronger alternative.
  3. Bring both outputs back into a third session. Ask the model to merge: "Here is a draft, here is a critique. Produce the revised version that incorporates valid critique and rejects invalid critique — explain each call."
  4. Read the explanations. Where the model rejects critique without a strong reason, you have your real edits.
  5. Optional: run step 2 a second time against a different model family (e.g. Claude → Gemini) to break shared blind spots.
  6. Lock the merged version as the spec and only then ask for execution.

Example prompt (step 2):

"A colleague produced the draft below in response to the task above. They asked for a brutal review before they ship. List every factual error, missing edge case, weak argument, and stronger alternative framing you can identify. Be specific. Do not soften."

Common failure + fix: the critique session is too polite — the model defends the original because it recognises its own voice. Fix: in step 2, never tell the model the draft came from itself, and explicitly instruct "your only job is to find problems; do not list strengths."

New AI Tools

Qwen 3.7 (Alibaba). The next major Qwen release, surfaced in the AISearch weekly roundup alongside DeepMind Co-scientist and ByteDance Lance. No detailed benchmarks in the digest yet; the relevance is competitive cadence — Alibaba shipping a numbered point release inside the same week as Google I/O, NVIDIA Polar, and Together OSCAR means the open-weights Chinese frontier is now setting its own schedule, not reacting to US labs. Source: AISearch.

DeepMind Co-scientist. A research-agent product from Google DeepMind designed to act as a hypothesis-generating, literature-searching collaborator for working scientists. Audience: research orgs, R&D leaders in pharma and materials, and consulting firms building scientific-knowledge agents. Source: AISearch.

NVIDIA Polar (open-source, NVIDIA-NeMo/ProRL-Agent-Server). Reinforcement-learning framework that proxies LLM calls during rollouts to capture token-faithful trajectories on unmodified coding harnesses (Codex, Claude Code, Qwen Code). Audience: any team training open coding agents — the +22.6-point SWE-Bench gain on Qwen3.5-4B is the headline number this week. Source: arXiv.

AI Personality of the Week

Andrej Karpathy. Karpathy joining Anthropic is the personnel move of the year, not the week. He was OpenAI's founding research scientist, ran Tesla Autopilot, returned to OpenAI in 2023, left again in early 2024 to run Eureka Labs as a one-person AI-education operation, and has spent the intervening two years effectively serving as the field's most influential public teacher. Moving him onto Anthropic's payroll concentrates technical credibility, public-explainer authority, and recruiting gravity in a single lab at exactly the moment that lab is posting its first profitable quarter and co-presenting an encyclical with the Pope. The second-order effect to watch: how Karpathy reorganises Anthropic's external research communication — historically a strength, but one that compounds further with a teacher of his profile inside the building. Source: Creators' AI weekly digest.

Catch-All

"How helping your rivals makes you harder to beat." Big Think Business ran an Eric Markowitz essay arguing that the firms most resistant to disruption are the ones that openly help their competitors — citing patterns from Toyota's open-sourcing of patent libraries, Tesla's 2014 patent release, and historical cases from the steel and shipping industries. The relevance for AI in W22 specifically: Anthropic's open interpretability work, NVIDIA open-sourcing Polar, and Together open-sourcing OSCAR all fit the same pattern. The argument is that publishing tools your competitors will adopt is a moat strategy, not a generosity strategy. Source: Big Think Business.

Show Notes (bullets only)

  • Andrej Karpathy joins Anthropic — the highest-profile transfer in industry history.
  • Anthropic on track for first-ever profitable quarter; reframes the "frontier labs are bonfires of cash" narrative.
  • Pope Leo XIV publishes Vatican's first AI encyclical, co-presented with Anthropic's Christopher Olah.
  • Claude Opus 4.8 ("Mythos") rolling out to all Anthropic customers.
  • Google I/O 2026 ships 100 announcements across Gemini, AI Mode, Android XR, Workspace, and Vertex.
  • Sequoia: the next $1T company won't sell software — it will sell the work.
  • Big Four enterprise AI agent playbook formalises; consulting-led agent delivery is now the dominant enterprise pattern.
  • NVIDIA Polar improves Qwen3.5-4B by 22.6 points on SWE-Bench Verified via token-faithful RL.
  • Together AI's OSCAR cuts KV cache memory 8× at near-BF16 accuracy.
  • SK Hynix becomes first memory company to cross $1T market cap.
  • SpaceX S-1 filing reveals xAI's standalone finances are materially uglier than the headline valuation implies.
  • Meta cuts 8,000 staff in its largest round since 2023.
  • Qwen 3.7, DeepMind Co-scientist, ByteDance Lance, Mega ASR, and a LeRobot humanoid all ship in the same week.

Weekly Patterns (Inference)

  1. Inference Anthropic is consolidating four kinds of legitimacy at once — technical (Karpathy, Opus 4.8), financial (profitability), moral (papal encyclical), and political (Olah at the Vatican podium). No competitor has all four right now, and the gap is visible week-over-week.
  2. Inference The frontier lab race is bifurcating on profitability. Anthropic's reported profitable quarter, paired with the SpaceX filing showing xAI's ugly standalone economics, means investors can now distinguish between "scaled and on a credible path" and "scaled and still bleeding" — that distinction did not exist in 2025.
  3. Inference Sequoia's "sell the work, not the software" thesis and the Big Four's agent factories are the same story from two angles. The buyer doesn't want a tool; the buyer wants the deliverable. The companies positioned to capture this are services-shaped, not SaaS-shaped — and the implications for incumbent SaaS valuations are material.
  4. Inference Inference-stack efficiency is now a weekly cadence. OSCAR (8× KV memory), EAGLE 3.1 (deeper speculative decoding), and Polar (cheaper agent RL) all landed in one week. Serving cost is dropping faster than model quality is plateauing, which means the unit economics of agent products improve every quarter without product-level changes.
  5. Inference Open-source publication is becoming a moat play, not a moat surrender. NVIDIA open-sourcing Polar, Together open-sourcing OSCAR, and Anthropic's open interpretability publishing all share a pattern: the publisher commoditises a layer below them, locking the value to a layer they control. The Big Think "help your rivals" essay this week formalises what the labs are already doing.
  6. Inference The Chinese frontier is on its own schedule. Qwen 3.7 shipping in the same week as Google I/O, with no apparent gating against US announcements, suggests Alibaba (and by extension ByteDance and DeepSeek) are no longer pacing themselves against US releases. That's a competitive posture shift worth tracking.
  7. Inference Memory is the binding constraint on frontier compute, not logic. SK Hynix's $1T cap is the cleanest market evidence; OSCAR's existence is the cleanest technical evidence. AI buyers should expect HBM scarcity, not GPU scarcity, to drive 2026-H2 pricing and contract terms.
  8. Inference The labour-market story is moving from speculation to data. Anthropic's exposure report this week is the highest-signal labour-impact analysis published by any frontier lab. Expect HR, workforce-planning, and consulting firms to treat the methodology — task-level Claude-exposure scoring — as the new template.