The frontier labs are releasing big capabilities and publishing thought pieces on the future, not just of the industry but for all of humanity. We’re seeing step changes in model capabilities, advances in memory and personal context, concerning observations in cybersecurity, and tangible progress on life sciences research. These capabilities have real impact on the future of work. Meanwhile, policy lags behind technological progress, despite the need for governance to ensure that the benefits of those advances diffuse to all.
RELEASES
① Dreaming: Better memory for a more helpful ChatGPT
Summary: OpenAI is rolling out “dreaming”, a new memory architecture for ChatGPT that replaces the previous saved memories system with a background process that continuously synthesizes context from chat history. Unlike the prior approach of explicit memory saves, dreaming automatically curates and updates memories over time with three design goals: 1/ carry forward useful context, 2/ follow preferences and constraints, and 3/ stay current over time.
Signal: Persistent and context-aware memory is becoming table stakes for frontier AI products, but the engineering challenge is serving it at scale across hundreds of millions of customers. OpenAI’s 5x compute reduction to serve dreaming signals that personalization is crossing over from a premium feature to core infrastructure. The three design goals offer a practical blueprint for designing a personalization layer that doesn’t go stale.
② Claude Fable 5 and Claude Mythos 5
Summary: Anthropic launched Claude Fable 5, a Mythos-class model with capabilities exceeding any prior generally available models, available for general use at $10/$50 per million input/output tokens. Fable 5 includes safeguards that route queries on biology, chemistry, cybersecurity, and distillation to Claude Opus 4.8. Anthropic also released Claude Mythos 5 to Project Glasswing partners with safeguards lifted for cybersecurity and soon to select biology researchers with biology and chemistry safeguards lifted. Both models can work autonomously for longer periods than previous Claude models but also require 30-day data retention for safety monitoring purposes, a shift that matters for organizations bound by zero-data retention policies.
Signal: Anthropic is making explicit the tradeoff between releasing frontier capabilities and managing misuse risk. Fable 5 uses an AI classifier to prevent jailbreaks, redirecting sensitive queries rather than refusing them outright, maintaining usability while enforcing safety. This approach offers a pattern for balancing capability access with risk containment as models become more powerful.
③ It’s safe to close your laptop now: Hosting coding agents on Amazon Bedrock AgentCore
Summary: AgentCore Runtime solves the constraint that coding agents require an always-on developer machine, giving each agent session an isolated Firecracker microVM with a persistent workspace that is retained for 14 days of inactivity, an interactive PTY-backed shell, and deterministic command execution via HTTP/2. Runtime can host Claude Code, Codex, Kiro, OpenCode, Cursor CLI, Gemini CLI, and custom harnesses. Each agent can work with any model via Bedrock, direct provider APIs, or custom LLM gateways. Multiple agents can run in parallel against the same codebase without localhost port collisions, shared credentials, or exposed SSH keys.
Signal: Coding agents were previously bound to developer laptops by technical necessity. AgentCore Runtime moves that constraint from a developer’s machine to the infrastructure layer, enabling parallel execution, enterprise identity and audit controls, and persistent state across sessions. Managed environments for securely scaling agent workloads is a key enabler for enterprise AI adoption.
THOUGHT LEADERSHIP
① When AI builds itself
Summary: Anthropic published internal data showing that, as of May 2026, more than 80% of code merged into its codebase was authored by Claude, with engineers shipping 8x as much code per day as they were in 2024. The length of tasks AI systems can reliably complete on their own has been doubling every four months. Claude still falls short on autonomous goal-setting, i.e., choosing which problems to work on in engineering and research, but Anthropic’s data shows that gap is narrowing. They outline three possible futures ranging from stalled progress with widely diffused capabilities to full recursive self-improvement.
Signal: Human review of AI-generated work is already becoming the binding constraint, i.e., if humans can’t audit code as fast as Claude generates it, oversight breaks down before capability does. Anthropic says it would support a verifiable pause in frontier AI development if one could be coordinated; prior verification regimes have been built as precedent, but those took decades to build. Anthropic says we don’t have that long.
② Built to benefit everyone: our plan
Summary: OpenAI published a strategic plan with three goals: 1/ build an AI system capable of automated research, 2/ accelerate economic growth through AI adoption, and 3/ give everyone access to personal AGI. OpenAI believes that, by March 2028, a significant fraction of its research will be done by AI systems in tandem with their researchers. They frame this as a transition from developing frontier capabilities to ensuring that AI is affordable, accessible, and safe enough for all to benefit.
Signal: OpenAI argues that transformative technology either concentrates power or distributes it, comparing AI to electricity’s historical impact. The company contends that broad access, open ecosystems, and public coordination determine whether AI’s benefits concentrate or diffuse. This frames broad distribution as a precondition for AI benefiting humanity.
③ Policy on the AI Exponential
Summary: Anthropic published two policy frameworks to address AI advancement outpacing policymaking: 1/ an advanced AI framework, calling for governance of increasingly capable systems, 2/ an economic policy framework, addressing the distribution of financial and capability benefits. The advanced framework applies to models trained with more than 10^25 FLOPS or by companies earning over $500M in AI revenue or spending over $1B on AI R&D. It requires frontier developers to meet four requirements: 1/ transparency, 2/ independent evaluation, 3/ security, and 4/ regulatory authority.
Signal: Anthropic proposes governmental authority to block frontier AI deployment that poses catastrophic risk, addressing four categories: 1/ biological, 2/ cyber, 3/ loss of control, and 4/ automated R&D. With accelerating AI capability, this framework shifts deployment authority from company discretion to requiring government approval for high-capability systems.
SECURITY
① Zero Trust for AI agents
Summary: Anthropic presents a tiered Zero Trust security framework for deploying AI agents in the enterprise. They cover security considerations unique to agentic systems, the current threat landscape for agents, and an eight-phase implementation plan. They link a deep dive that they authored, providing an accessible perspective for technology leaders and concrete guidance for security practitioners.
Signal: Frontier AI is compressing the time from vulnerability to working exploit from months to hours, and agentic attackers compound this by operating with unlimited patience at near-zero per-attempt cost. The relevant test for any defensive control is no longer whether it raises the difficulty bar, but whether it makes the attack impossible. Security teams that don’t account for this asymmetry between attackers and defenders will be designing defenses for a threat model that is no longer relevant.
② What we learned mapping a year’s worth of AI-enabled cyber threats
Summary: Anthropic analyzed 832 accounts banned for malicious cyber activity between March 2025 and March 2026. They found three things: 1/ the share of threat actors classified as medium risk or higher jumped from 33% to 56% over that period; attackers were also shifting from gaining initial access to post-compromise activity, 2/ cyber attacks are increasingly autonomous, i.e., higher risk actors design architectures that chain discrete attack stages together with minimal human input, and 3/ the MITRE ATT&CK framework doesn’t yet capture this type of attack that uses agentic orchestration.
Signal: Traditional risk signals, e.g., technique count, platform choice, no longer reliably identify high risk actors, because AI lets less sophisticated attackers perform post-compromise work that previously required deep expertise. MITRE ATT&CK doesn’t yet cover agentic orchestration, meaning AI-enabled attackers may be systematically under-classified by detection tooling. This gap should be treated as a known blind spot, and detection rules should be revisited.
RESEARCH
① Making Claude a chemist
Summary: Anthropic benchmarked three Claude models against ChemDraw and MestReNova, two dedicated chemistry software tools, on NMR spectroscopy, a technique that chemists use to determine the structure of molecules. Using 20 compounds whose results were posted after the models’ training cutoff, Opus 4.7 matched or outperformed both tools at predicting what a spectrum should look like for a given molecule and could also work the problem in reverse, i.e., proposing a molecular structure from a spectrum alone, a task existing software typically leaves to the chemist.
Signal: NMR assignment is one of the most time-consuming routine tasks in synthetic chemistry, and this is one of the first demonstrations of a general-purpose model matching domain-specific scientific software without chemistry-specific fine-tuning. For technology leaders building AI-assisted research tools, it suggests that frontier model capability improvements are compressing the time to usefulness in specialized fields faster than domain-specific fine-tuning approaches.
② Introducing new capabilities to GPT‑Rosalind
Summary: OpenAI updated GPT-Rosalind, a model series built specifically for drug discovery and biological research, adding stronger reasoning for tasks like analyzing genetic data, designing drug candidates, and assisting with lab experiments. They also built four new benchmarks to measure performance on real scientific workflows, where the updated model outperforms the general-purpose GPT-5.5 on each. GPT-Rosalind is now available in research preview to eligible organizations globally, with Novo Nordisk announced as an early partner.
Signal: OpenAI built a domain-specialized model for life sciences, while Anthropic showed a general-purpose model could match dedicated chemistry software without any fine-tuning. OpenAI and Anthropic are taking opposite approaches to scientific research, and the outcome will shape how organizations build AI into research workflows, i.e., one towards domain-specific fine-tuned models and another toward general-purpose platforms.
③ Paving the way for agents in biology
Summary: Anthropic used their benchmark, VirBench, which includes 120 realistic viral sequences spanning 40 pathogens, to better understand the challenge of bridging agents to databases. The strongest models did not consistently achieve the level of accuracy required for reliable dataset construction, ranging from 16.9% to 91.3%, depending on model capability. However, when adding a deterministic retrieval layer, accuracy peaked at 99.7% while run-to-run variability also collapsed.
Signal: Frontier labs are racing to improve and scale model capability, but this research suggests that the real bottleneck in scientific research agents is in the data access layer. Well-designed deterministic tools and machine-readable data can matter more to accuracy and reliability than the model that is selected.
④ Measuring the impact of learning with AI in Sierra Leone and beyond
Summary: Google DeepMind conducted a randomized control trial with 1,763 junior secondary students across 12 schools in Sierra Leone’s Port Loko District over eight weeks, measuring the impact of Guided Learning in Gemini on math achievement. Students using Guided Learning gained +0.258 standard deviations, equivalent to 1.2-1.7 years of typical learning progress. Students in classrooms where teachers incorporated Gemini into roughly half their lessons saw larger gains of roughly 1.8-2.5 years of progress.
Signal: With widespread concern about generative AI becoming a shortcut for students, the trial shows that pedagogically grounded AI design can improve educational outcomes. However, students who entered with stronger math skills benefited the most, highlighting an achievement gap that remains unresolved with this approach.
CRITIQUES
① Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them
Summary: Gerben Wierda documents building a small project with Claude Code and concludes that subscription pricing ($180/month, per his own calculations) represented a 2.5x subsidy over what the same workload would have cost at API rates ($450/month). He argues that per-token pricing is the wrong unit for agentic workloads and suggests instead that per-task costs are more relevant. His tentative conclusion is that the current economics are unsustainable, and that the frontier labs may be absorbing significant losses on complex agentic use cases.
Signal: Per-task cost rather than per-token cost is the operative measure for agentic workloads. Given the subsidy levels and ongoing changes in inference pricing, subscription pricing and frontier lab subsidies are unlikely to hold. Current cost assumptions should be stress-tested, evaluating current inference usage against the possibility that pricing floors rise as labs move toward sustainable unit economics.
PAPERS
① How AI Agents Reshape Knowledge Work: Autonomy, Efficiency, and Scope
Summary: Frontier AI systems are shifting from conversational assistants to autonomous agents that are able to execute complex tasks end to end. Their ability to execute multi-step work changes how knowledge work gets done, e.g., agents perform approximately 48x more machine work per session, cut task time by 87%, reduce cost by 94%, and reduce user dissatisfaction by 55%.
Signal: As execution costs drop, both the pace and nature of work changes. The shift is from efficiency AI, i.e., doing the same work with fewer people, to opportunity AI, i.e., attempting work that was previously impractical. This is a structural change in what a small team can produce, which then has downstream implications for organizational design and hiring.
Note: Given the length of the paper, I used AI to summarize and help me more quickly understand the key takeaways of the paper.
Note: I heard the comparison of efficiency AI versus opportunity AI first from Nathaniel Whittemore.
Always be learning.
heeki reads #4
Written by Heeki Park, Principal SA @ AWS. Opinions are my own.
Alcurio is where alchemy meets curiosity.