Personal AI News Digest
6 topics from 3 sources
1,664 words · ~8 min read
Anthropic released Claude Haiku 5.5, the first Haiku update in about a year, priced to match OpenAI's GPT-6 Luna. On the Artificial Analysis Intelligence Index at max effort, Haiku 5.5 scores 43 (up 26 points from Haiku 4.5), ahead of GPT-6 Luna (38), Gemini 3.8 Flash (41), and GLM-5.3 Flash (42), but trailing Sonnet 5.5 (56) by 13 points. On Terminal-Bench 4.0, Haiku 5.5 scores 33% (up from 0% for Haiku 4.5), level with GLM-5.3 Flash and ahead of Gemini 3.8 Flash (20%) and Luna (13%). On AA-Briefcase (private agentic knowledge-work eval), it scores 1578 Elo, ahead of Kimi K3 and GLM-5.3, comparable to Muse Spark 1.3 at max. AutomationBench-AA shows 35% (versus 53–60% for Luna, Gemini 3.8 Flash, GLM-5.3 Flash), though a pre-release safety bug caused over-refusal; Anthropic is fixing this and expects the score to rise. Haiku 5.5 shows 10–50% gains in knowledge work and reasoning, beats Luna on most of these, and has 1M context with roughly 12k max output tokens.
Pricing: under 100K tokens $0.10/$0.50 per 1M input/output tokens with $0.01 cache reads per 1M; over 100K tokens $0.50/$2.50 with $0.05 cache reads per 1M. Anthropic claims it costs about 75% less to run than Haiku 4.5 on average. The model is the first Haiku with effort settings and adaptive thinking. Main caveat: at max effort it uses about 162k output tokens per Index task, roughly 3x GPT-6 Luna's ~50k; at equal score (38), Haiku 5.5 at high effort uses ~55k tokens versus Luna at max using ~50k.
Positioning: Anthropic positions Haiku 5.5 as a subagent paired with Opus 5.5 or Sonnet 5.5 for high-volume, cost-sensitive work such as summaries, compactions, and database queries. It launched live on Claude Platform, Claude Code, and in partner integrations: Cursor claims "10x less than Haiku 4.5" on shorter requests; GitHub Copilot in VS Code "matched Claude Sonnet 5 on many coding tasks while using fewer tokens and steps"; Devin reports 58.4% on FrontierCode 1.1 (ahead of Sonnet 5 at roughly one-eighth the cost per task) and recommends it as a "sidekick" under Opus 5.5 in Fusion. Added to Agent Arena, Code Arena WebDev, Text, Document and Vision arenas same day.
Sources AINews
Links [AINews] Claude Haiku 5.5 — better than GPT-6 Luna at the same pricing, shipped Haiku 4.5, @claudeai, @AnthropicAI, @ClaudeDevs, @ClaudeDevs, @ClaudeDevs, quickstart, @cursor_ai, @cursor_ai, @code, @cognition, @arena, @ArtificialAnlys, per-eval breakdown, @alexalbert__, @TheRundownAI, @LoganMarkewich, @jerryjliu0, @simonw, comparison, @AI_Screening
Anthropic cut Sonnet 5.5 cache-read pricing from $0.20 to $0.10 per 1M tokens (50% reduction), making Sonnet 5.5 about 20% cheaper on most long-running or agentic work. Simultaneously, Anthropic introduced tiered monthly Claude Platform API credits for subscribers: Max 5x ($100), Max 20x ($200), and Team (up to $500, pooled). These credits work on any model, including Haiku 5.5, and in third-party harnesses.
Sources AINews
Links @claudeai, @ClaudeDevs
Anthropic built computer-use and browser-use toolsets directly into the Python and TypeScript Claude SDKs. The SDK now runs the action loop and sends clicks and keystrokes to drivers from browser_use, Browserbase, E2B, or Daytona, eliminating the need for developers to write that loop themselves.
Sources AINews
Links @ClaudeDevs, quickstart
Kubernetes co-creators Craig McLuckie and Joe Beda are building Mecatl, an open-source cloud-native harness for coding agents that separates the agent loop from tool execution, state management, and execution environments. Unlike desktop-first harnesses that couple the loop, local execution, and session state in a single process, Mecatl is architected from the ground up for cloud deployment, keeping the agent loop independent of the client, model provider, state store, and execution environment. Stacklok also built ToolHive, an open-source platform for running and governing MCP servers at enterprise scale, evolved from a Docker-based desktop tool into a Kubernetes-based gateway with associated services including a registry and operator helper for managing MCP servers. ToolHive handles input-output authentication and enables enterprises already running Kubernetes to bring agentic workloads into those environments. Stacklok built an LLM gateway (AI Gateway) to help regulated enterprises—primarily banks, semiconductor companies, and telcos—control access, enforce budgets, manage reporting, and route requests across LLM providers. Unlike Glean's approach, Stacklok's gateway does not currently perform semantic model selection based on task; instead, model choice should happen in the harness where there is more context. The AI Gateway remains the company's primary closed-source product, though Beda indicated it is on the roadmap for open-sourcing. Stacklok's commercial offering is an enterprise control plane that provides consistent identity, authorization, policy, and auditing across its open-source projects (Mecatl and ToolHive). While the open-source components are designed to be independently useful for small teams, scaling across multiple clusters and clouds introduces new operational problems that the commercial product solves. Stacklok, which raised $17.5 million Series A in 2023 from Accel, Madrona, and Bain Capital, initially focused on software supply chain security before pivoting to Kubernetes-based agentic solutions.
Sources Latent.Space
Links Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?, Craig McLuckie, Joe Beda, Stacklok, raised a $17.5 million Series A, open source project on GitHub, via Mecatl GitHub, via GitHub, Stacklok’s “AI Gateway”, Glean’s solution
Latent.Space podcast episode featuring Periodic Labs founders Liam Fedus and Ekin Dogus Cubuk on building AI systems for autonomous materials discovery. The core thesis: intelligence alone is insufficient; new knowledge requires testing conjectures against physical reality. Periodic combines reinforcement learning grounded in physical experiments, AI-powered materials characterization, density functional theory simulations, and high-throughput labs to compress decades of scientific trial-and-error into months.
Key technical insights: (1) Materials science differs fundamentally from math and coding—it involves noisy, incomplete experimental data, stochastic labeling, and measurement uncertainty that demands decision-making under uncertainty and sample efficiency. (2) The materials discovery loop has three stages: predicting stable atomic configurations with desired properties, determining synthesis conditions to create them, and characterizing the result via X-ray diffraction and multimodal measurements. (3) Reinforcement learning environments are constructed around characterization tasks (e.g., phase identification from XRD patterns) rather than waiting weeks for final outcomes; this enables faster iteration and avoids models memorizing known answers. (4) Density functional theory (DFT) is powerful but limited—it cannot capture microstructure, superconducting temperatures, or strongly correlated systems; experiments remain ground truth. (5) Negative results and failed experiments are valuable training data, especially when published with full process lineage (conversations, intuitions, lab runs, computations), enabling models to learn the process of science rather than only its published results.
Lab automation strategy: Every instrument receives AI-driven intelligence to maximize data quality and reduce noise floor. Latency and cost constraints mean some reasoning must run on-device; tool calls to DFT or other simulations can dominate inference time. Automation targets routine, time-consuming tasks (e.g., characterization bottlenecks) rather than dexterous work; humans remain essential for high-level campaign direction and scientific intuition. Robots reduce mistakes and enable AI systems to detect anomalies (e.g., cyclic loading errors) across longitudinal data.
Model and compute strategy: Periodic uses a mix of open-source and closed-source models, fine-tuned on proprietary experimental data. Access to high-quality, domain-specific data enables compute efficiency gains over frontier models; this data advantage is irreplaceable. Training on full campaign traces—not just final outputs—allows models to learn reasoning strategies for physical-world discovery. Even frontier models will require experiments; scientific discovery is by definition what models haven't been trained on.
Team composition reflects the irreducibly multidisciplinary nature: AI researchers, computational materials scientists (DFT, force fields), experimentalists, hardware engineers, and LLM experts. Leadership includes Dima Bahdanau (inventor of neural attention), Joe Checkelsky (MIT physicist, superconductivity expert), Muratan Aykol (computational materials science), and Daniel Chica (synthesis expert from Mercury Canasides' group). The model mirrors Bell Labs: experienced researchers do hands-on work, not just management.
Monetization and commercialization: Forward-deployed engineering teams embed at semiconductor and materials partners (e.g., Taiwan fabs) to integrate Periodic's AI and computational tools on-site, training models on partner data so they own their own intelligence. As autonomy grows, pricing can shift from services to outcomes. The goal is to make materials science and solid-state physics commercially viable, attracting talent and capital to the field.
Path to superconductors: Rather than waiting for theory of high-temperature superconductivity, Periodic emphasizes synthesis superintelligence to explore chemical spaces (e.g., transition metal oxides, nickelates, cobaltates). Historical precedent: a Japanese group discovered MgB₂ (highest-temperature ambient-pressure conventional superconductor) by trying ~30,000 materials; Periodic's automation could compress that search into weeks, dramatically increasing the surface area for discovery.
Open-source contributions: Periodic maintains active contributions to PyTorch, Materials Project, JAX-MD, TorchSim, and LLM infrastructure (XLANG, Megatron). Academic grant program funds university groups advancing synthesis superintelligence.
Sources Latent.Space
Links Synthesis Superintelligence: …, Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic…, Periodic
Sam Newman, author of Building Microservices and the forthcoming Building Resilient Distributed Systems, discusses microservices as an architecture of last resort, distributed systems design, and AI's role in software development. He was in the room when the term "microservices" was coined in the early 2010s at an architecture symposium; James Lewis and Martin Fowler published the first definition in March 2014, and Newman's book followed in 2015.
Newman distills the eight fallacies of distributed computing into three core rules that cause most outages: (1) information takes time to travel (speed-of-light constraint), (2) resources go offline, and (3) resources are finite—with #3 causing most real-world distributed systems outages. He emphasizes observability as essential and frames resilience decisions around four military system outcomes: graceful degradation, robust recovery, rapid recovery, and graceful surrender.
On idempotency, Newman discusses two approaches: idempotency keys and fingerprints. He stresses that business context must inform whether systems fail open or closed on errors. Independent deployment and team autonomy are central to microservices adoption; many teams misunderstand these requirements.
Regarding AI, Newman warns against "cognitive surrender"—using AI in ways that increase context switching and reduce critical thinking rather than freeing engineers from drudgery. He advocates hedging vendor and model choices to avoid lock-in, and notes that deterministic code can often replace LLM functionality at lower cost. He cautions that much of the tech world remains naive about LLM capabilities and limitations. When writing his new book, he used NotebookLM for research but manually verified every link it surfaced.
Newman recommends starting with modular architecture to encourage better system design, and emphasizes that "production is truth." He suggests reading Fundamentals of Software Architecture, Balancing Coupling in Software Design, Learning Domain-Driven Design, David L. Parnas's Software Fundamentals, and Just Enough Software Architecture.
Sources The Pragmatic Engineer
Links Building resilient systems with Sam Newman, YouTube, Apple, Spotify, Building Microservices, Building Resilient Distributed Systems, Microservices: a definition of this new architectural term, eight fallacies of distributed computing, by David Woods, Fundamentals of Software Architecture, Balanced Coupling, On the criteria to be used in decomposing systems into modules, An architecturally-evident coding style: making your design visible, Designing Data-intensive Applications with Martin Kleppmann, chapters, https://x.com/samnewman, https://www.linkedin.com/in/samnewman, https://www.amazon.com/dp/1492034029, https://martinfowler.com/microservices, https://samnewman.io/books/building-resilient-distributed-systems, https://margaretstorey.com/blog/2026/02/09/cognitive-debt, https://www.amazon.com/dp/1098175514, https://www.amazon.com/Balancing-Coupling-Software-Design-Addison-wesley/dp/013…, https://www.amazon.com/Learning-Domain-Driven-Design-Vlad-Khononov-ebook/dp/B09…, https://www.abebooks.com/9780201703696/Software-Fundamentals-Collected-Papers-D…, https://www.georgefairbanks.com/book