The AI Enthusiast.

Owen Templin — History Retold, and a workflow built with AI. #

Owen is presenting at this week's meeting. Before he does, go watch the channel — this is genuinely good work and it deserves an audience in the room.

@HistoryRetoldAnimated — 75 videos, long-form animated history, typically 9–16 minutes. The channel's own description is the best pitch anyone could write for it:

"History isn't boring. It was just told wrong. Dry textbooks and forgotten dates buried the most dramatic stories ever lived — assassinations in cathedrals, empires built by bankers, armies that shouldn't have won and did anyway. History Retold cuts through all of that and gets to what actually matters: the turning points, the figures, the decisions that changed everything. No filler. Just history, the way it was always meant to be told."

That's the whole thesis, and the catalog delivers on it. Seventy-five animated episodes, all produced solo. If you have ever tried to make a single ten-minute video, you already know what that number means.

Why it's worth your time

Owen uses AI as part of how he actually makes these videos. That's the reason he's a good fit for this group: his workflow shows AI being used in ways most of us haven't thought to try, on a real project with finished results you can watch.

Anthropic gave Claude hands — and a 99.3% success rate on quantum hardware. #

Announced August 27, 2026

Anthropic previewed the Model Hardware Standard (MHS), a protocol that lets AI agents drive physical machinery: microscopes, liquid handlers, robotic arms, laser rigs. It is deliberately the MCP playbook applied to atoms instead of software — model-agnostic, initially a research preview for a hand-picked set of science, robotics and manufacturing organizations, with a stated intent to eventually open-source it so any device manufacturer can adopt it.

The pitch is integration time. Standing up and wiring together instruments in a lab or plant typically takes weeks or months, and most devices never talk to each other at all. Anthropic claims MHS collapses that to hours or minutes, and it can operate multiple instruments in parallel. Development started as a collaboration with HHMI Janelia Research Campus.

The launch-partner numbers are the interesting part, because they're specific rather than aspirational. Genentech automated a protein assay procedure. Carnegie Mellon ran drug-discovery experiments roughly 3× faster. QuEra Computing had an agent recover a laser's precise operating frequency with a 99.3% success rate. That last one matters — it's an agent doing closed-loop calibration on quantum hardware, not summarizing a PDF.

Claude stopped forgetting — and started showing you its notes. #

Announced August 25, 2026

Anthropic merged the memory systems behind Claude chat and Claude Cowork into one. What Claude learns in a Cowork session it knows in chat, and vice versa. The stated target is the single most tedious thing about working with agents — re-briefing the assistant on context it already had.

Two mechanical changes sit underneath the headline. First, memories are now written on the fly during a conversation rather than assembled in a post-hoc review at the end of the session. Second, memory is now inspectable and editable: everything Claude retains is listed by topic under Settings → Memory, where you can read, edit, or delete individual items. That's a meaningful shift from "the model remembers things about you and you hope it's right" to something you can audit.

By default Claude will not store personal or sensitive categories — health data, race, ethnicity, religious belief, politics, gender identity and similar — with an explicit opt-in toggle ("include sensitive topics in memory") for users who want it. The feature is on by default for Free, Pro and Max across web, desktop and mobile.

Qwen shipped a 2.4T flagship and a 27B that runs on your Mac. Only one of them is actually open. #

Released August 13–14, 2026 · Qwen3.8-Max and Qwen3.8-27B · Muse Code beta, Spark 1.2, Glimmer

First, a correction to how this was initially framed. Qwen 3.8's coding performance is a genuine strength, not a weakness. The headline software-engineering number is SWE-bench 87.3% — ahead of GPT-5.5 (82.6%) and behind Claude Opus 4.8 (89.2%), a two-point gap to the frontier rather than a rout. Add Frontend Code Arena #4 at 1,668 Elo and #2 among open-weight models on the Vals Index (66.1). Anyone who came away thinking Qwen 3.8 was weak at coding was reading a cherry-picked slice of the table.

Second, and more important: there are two models, and the distinction is the whole story.

Max is the first Max-class Qwen to get open weights, which is the historic bit. But the 27B is the one that matters to most people, and it's the reason for the community energy: its predecessor, Qwen3.6-27B, was one of the most-praised local models of the year — the kind people describe as punching above its weight class — and 3.8-27B inherits the new generation's training on hardware you already own.

Where it's actually weak. Two places. HLE 43.6 on the broad-knowledge frontier exam is last among the four flagships (Fable 5 leads at 53.3, GPT-5.6 Sol 47.2, Opus 4.8 45.7) — breadth of world knowledge is the real gap. And on SWE-bench Pro — real GitHub issues in professional codebases, a much harder test than Verified — Max scores 67.7 against Anthropic's frontier tier at 80.0–80.3. Even there it beats GPT-5.6 Sol (64.6). The 27B's own model card reports SWE-bench Pro 61.7 and OSWorld-Verified 84.3.

What the community actually says. Infrastructure moved fast — Baseten and Teknium confirmed rapid integration support, and the prevailing read is that "open models are at or near frontier parity on meaningful workloads." Two counterweights are worth carrying into the meeting. Jamin Ball's point that a >2T-parameter model needs at least 8 H100/B200 GPUs in a supernode, so "open weights" and "accessible" are not the same claim. And a licensing flag: users have noted geographic restrictions naming the USA, EU, UK and Korea — verify that against the actual license text before anyone builds on it. The sharpest technical skepticism isn't about the scores at all: it's whether the gains come from novel pretraining or from better post-training and harness design, which would make them easier for everyone else to copy.

Meanwhile Meta Superintelligence Labs reshaped the open-weights picture from the other end. The Llama line has effectively been succeeded by Muse: August brought Muse Code (beta) and Muse Spark 1.2 (weights under a modified Llama Community License), plus Muse Glimmer — a 29.6B dense multimodal model distilled from Muse Spark under Apache 2.0. Glimmer and Qwen3.8-27B now land in the same size class, on the same license, aimed at the same GPU.

Nvidia is reportedly buying the place every open model lives — for $12.9B. #

Reported August 26, 2026 (The Information) · talks first reported August 24 · status: unconfirmed, no signed contract

Nvidia has reportedly agreed to acquire Hugging Face for $12.9 billion. The Information broke it on August 26, with CNBC and Forbes following on August 27 (Forbes rounding to $13B). Handle this one carefully: neither company has confirmed publicly, and reporting is explicit that no signed contract exists yet — the talks can still collapse. Treat it as a well-sourced report, not a done deal.

The price tells the story of the last three years. Nvidia put $235 million into Hugging Face's Series D in August 2023, at a $4.5 billion valuation. A $12.9B purchase is roughly a 2.9× step-up in three years for a company founded in 2016 whose product is, essentially, the place developers put their model weights.

Why this matters more than the number. Read it against story 04. Qwen3.8-27B, Qwen3.8-Max and Meta's Muse Glimmer all shipped their weights to Hugging Face this month. It is the default distribution layer for open-weight AI — the thing that makes "we released the weights" a meaningful sentence. The buyer is the company that sells the GPUs those weights run on.

• Can the hub stay neutral? A vendor-neutral registry owned by the dominant silicon vendor is a genuine structural question. Every model card, download path and default inference recommendation sits inside that conflict.
• It's a shot at the hyperscalers. Own the chips and the discovery layer, and you can route developers toward your own stack before AWS, Azure or GCP get a look. Analysts are framing this as the deal's real rationale.

There's also a European angle worth naming: Hugging Face was French-founded, and its acquisition by a US chipmaker is landing badly in EU tech-sovereignty conversations.

DeepSeek open-sourced the agent loop itself — 95,000 stars in 48 hours. #

Published August 13, 2026 · MIT license · developer preview

DeepSeek's open-source agent harness — the CLI is dsh — passed 95,386 GitHub stars and 8,826 forks in roughly two days, one of the fastest adoption curves GitHub has recorded for a developer tool. It's MIT licensed, shipped as a developer preview, and DeepSeek explicitly warns that APIs and plugin contracts may change before a stable release.

The architecture is one idea taken to its logical end: everything is a plugin. Not just tools and models — the model adapter, tool registry, session log, sandbox, and the agent loop itself are all swappable. Models, tools, skills, sessions, sandboxes, storage, loops, scheduling and the UI are all recomposable pieces.

The design detail worth stealing regardless of whether you adopt it: an append-only session log that records everything the model saw — system prompts, reasoning, tool calls and results, subagent scheduling, and every context injection. Resume, fork, search and replay all operate on the same event stream. That is a real answer to the biggest operational problem with agent harnesses, which is that when a long run goes wrong you usually cannot reconstruct why.

Read together with story 04, the strategic picture is clear: Chinese labs are competing on the harness and the license, not just the weights. Anthropic's moat is Claude Code plus MCP plus enterprise controls; dsh attacks exactly that layer, for free, with the loop itself pluggable.

OpenAI says its next model did new math for $2,000. It also loses $1.22 for every dollar it earns. #

Astra reported August 2026 · confidential S-1 filed May 22, 2026

The most striking claim of the month: OpenAI says an internal version of its next major model — codenamed Astra — worked through ten previously unsolved problems in mathematics and theoretical computer science, including sphere-packing density, at a reported compute cost of roughly $2,000. Treat the specifics as company-reported until independently verified, but note what's being asserted: not "solved a hard benchmark," but produced new mathematics, cheaply.

Elsewhere OpenAI shipped GPT-Live full-duplex, where the voice model can interrupt and be interrupted naturally — the end of walkie-talkie voice AI. And ChatGPT crossed roughly 1 billion weekly active users.

The corporate storyline is less tidy. OpenAI filed a confidential draft S-1 with the SEC on May 22, 2026, without disclosing size, terms, timing, exchange or final valuation. Reported valuations cluster between $730B and $852B against annualized revenue past $25B. Reuters reported in late June that OpenAI may wait until 2027 to go public. Q1 2026 saw a loss of roughly $1.22 per dollar of revenue.

MCP and A2A now live under one roof — plus a primer on the one you don't know. #

A2A joined AAIF August 17, 2026 · new MCP roadmap published August 22, 2026

Two things happened inside the window, and together they're a bigger deal than any single spec change.

August 17 — the two protocols consolidated. The Agent2Agent (A2A) protocol officially joined the Agentic AI Foundation, the Linux Foundation directed fund that Anthropic donated MCP to. AAIF was co-founded by Anthropic, Block and OpenAI, with support from Google, Microsoft, AWS, Cloudflare and Bloomberg, and is anchored by MCP, goose and AGENTS.md. So the agent-to-tool protocol and the agent-to-agent protocol are now governed by the same neutral foundation, backed by essentially every major lab and cloud simultaneously. The "USB-C and TCP/IP of the agent era" framing stops being an analogy and starts being an org chart.

Worth knowing where A2A came from, because it's the same story as MCP. A2A is Google's protocol. Google announced it on April 9, 2025 at Google Cloud Next, then donated it to the Linux Foundation on June 23, 2025 at Open Source Summit North America — handing over the specification, the SDKs and the developer tooling, and seeding a vendor-neutral project launched with AWS, Cisco, Google, Microsoft, Salesforce, SAP and ServiceNow. It's now run by a Technical Steering Committee drawn from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow, with 150+ organizations backing it and production deployments across several industries.

So both halves of the agent stack followed an identical arc: one lab builds a protocol, gives it away to the Linux Foundation, and it becomes an industry standard. Anthropic did it with MCP, Google did it with A2A, and as of August 17 the two live under the same roof. Neither company controls the thing it created anymore — which is the point.

MCP vs. A2A — you already know one of these

Most of the room knows MCP and has never touched A2A. The distinction is genuinely simple, and the one-line version is: MCP gives an agent its toolbox; A2A lets that agent hire other agents.

August 22 — a new roadmap, and it's mostly about identity and security. Five priority areas, and the weighting tells you what the enterprise blockers actually are:

• Agentic messaging primitives — server-initiated events, maturing the Tasks extension, coordination on streaming and long-running workloads.
• HTTP-native transport unification — extending HTTP deployment to local servers via Streamable HTTP over stdio, so local and remote stop being different problems.
• Agent identity and enterprise-ready security — DPoP (Demonstrating Proof of Possession), Workload Identity Federation, and standardized agent delegation via token exchange.
• Improved primitives — standardized tool result handling, and progressive tool discovery to cut model overhead.
• SDK developer experience — ergonomics, spec conformance testing, documentation.

Item 3 is the one to read twice. DPoP, workload identity federation and delegated tokens are not agent research — they are the standard vocabulary of enterprise auth. That's a protocol growing up into procurement. Item 4's progressive tool discovery is the quietly practical one: if you've ever watched context get eaten by a hundred tool definitions the model never calls, that's the fix.

Background for both: the 2026-07-28 spec already made MCP stateless — sessions and initialization handshakes are gone, so servers scale horizontally without holding state — plus authorization hardening and a formal 12-month deprecation policy. Maintainers David Soria Parra and Den Delimarsky called it the most important release since remote MCP.

Anthropic is paying outsiders $5M to find out if Claude is bad for you. #

Announced August 25, 2026 · applications due September 21, 2026 · notifications October 5, 2026

Anthropic launched a $5 million grant program funding independent research into how AI affects user wellbeing. Grantees get direct funding, access to Claude models and technical guidance drawn from Anthropic's Safeguards team methodology — and they work independently, with a requirement to publish findings as open-source projects.

The gap it targets is specific and uncomfortable: nobody has good measurements for how models behave during sensitive conversations — emotional support, mental health crises and companionship-seeking interactions. Anthropic is explicitly recruiting outside its own discipline: clinicians, psychologists and methodologists rather than ML researchers.

The methodology bar is where this gets credible. Funded evaluations must define success metrics clearly, involve clinical experts in the design, test both safeguards and potential harms, reflect realistic multi-turn conversations, and validate graders against subject-matter experts. That last requirement — validating the grader, not just the model — is the thing most AI evaluation work skips.

Anthropic made money. Most of its customers still can't prove they did. #

Q2 results reported August 15, 2026

The vendor side finally posted a profit. Anthropic reported preliminary Q2 2026 revenue above $11.5 billion — roughly 14× the $787 million of the same quarter last year — with positive adjusted operating income, reported around $559 million, years ahead of its own schedule. Figures are preliminary and could move.

The customer side is a different chart. 79% of organizations report challenges adopting AI, a double-digit increase over 2025, and 54% of C-suite executives say AI adoption is tearing their company apart. Usage isn't the problem: 70% of employees and 94% of the C-suite use AI tools at least 30 minutes daily, 64% of executives two hours or more, and 88% of organizations use AI in at least one business function. But only about one third have begun scaling it enterprise-wide, and only 29% see significant ROI from generative AI — 23% from AI agents. AI "super-users" deliver up to 5× productivity gains with no mechanism to propagate their practices.

Developer discussion has converged on the same diagnosis from the bottom up. AI-generated code doesn't necessarily reduce maintenance and may increase it; companies are reportedly spending 44% of their tokens fixing bugs their own AI generated. Research found developers reported being faster while actually being slower, once you count finding errors, steering the agent and waiting. Uber's COO said AI spend hadn't produced a measurable increase in projects or productivity. The recurring conclusion: generation speed is no longer the bottleneck — verification capacity is. Agents are useful for bounded tasks, not autonomous software production.

Getting more out of Claude this month — mostly by deleting things. #

Drawn from August's releases and Anthropic's own prompting documentation for Opus 5 and Fable 5. The through-line is counterintuitive enough to lead with: the Claude 5 models need fewer instructions, not more. Rules written to protect against older models now cost you quality.

Commands worth putting in your muscle memory

Claude Code now exposes 90+ slash commands, and the built-ins are designed to be a template — writing your own skill is markdown in a file, portable across models.

Prompting Opus 5 (and Opus generally)

• Delete your verification instructions. Opus 5 verifies its own work unprompted. Instructions like "include a final verification step" or "use a subagent to verify" cause over-verification — removing them cuts wasted tokens with no quality loss. Same for "double-check your answer."
• Effort is your cost dial, not your verbosity dial. Default is high; use low and medium liberally where quality holds, xhigh for demanding agentic work. Lowering effort reduces thinking, not visible response length — if you want shorter answers, ask for shorter answers.
• Written files run long. Add explicit length calibration: "Match the length of written documents to what the task needs; do not pad with filler sections, redundant summaries, or boilerplate."
• Never tell a reviewer to be conservative. If your review prompt says "only report high-severity issues," Opus 5 obeys literally and reports less. Ask for everything, filter in a separate pass.
• Constrain scope on narrow tasks. Opus 5 will expand a task and apply its own judgment about what it should be. Say so explicitly when you want exactly what you asked for.
• Cap delegation if cost matters. CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH, CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS and the SDK's max_budget_usd are the deterministic controls (requires Claude Code 2.1.217+).
• Keep thinking on. Thinking-disabled mode can leak tool calls as plain text or emit internal XML tags. Thinking on at low effort beats thinking off at similar cost.
• Context window is 1M tokens as both default and maximum, with instruction-following holding throughout.

Prompting Fable 5

• Point it at your hardest unsolved problem. Anthropic's explicit guidance: teams testing Fable 5 only on simpler workloads undersell it. It's built for work that takes a person hours, days or weeks.
• Budget for long turns. Individual requests can run many minutes at high effort; autonomous runs can go hours. Adjust client timeouts and progress indicators, and prefer checking on runs asynchronously over blocking.
• Do not ask it to explain its reasoning. Prompts or skills that tell the model to echo, transcribe or narrate its internal reasoning can trigger the reasoning_extraction refusal and cause elevated fallbacks to Opus 4.8. Read structured thinking blocks instead. Audit your existing skills for "show your work" language before migrating.
• Ground progress claims. Nearly a magic phrase — in Anthropic's testing it nearly eliminated fabricated status reports: "Before reporting progress, audit each claim against a tool result from this session. Only report work you can point to evidence for."
• Give it a memory file. Fable 5 performs notably better when it can record lessons from prior runs and reference them. One lesson per file, one-line summary at the top, delete what turns out wrong.
• Give the reason, not just the request. "I'm working on [larger task] for [who]. They need [what the output enables]. With that in mind: [request]."
• Refactor old skills. Skills written for prior models are often too prescriptive for Fable 5 and actively degrade output.
• Know the guardrails. Fable 5 runs safety classifiers on offensive cybersecurity and biology/life-sciences content, and benign work in those areas can trip them. Configure fallback to Opus 4.8 if you work near those domains.

Four verticals — for the parts of the audience whose day job isn't building models. #