Quick Hits

• Open-weight releases crowded the paid API stack: @FishAudio launched Drama 3, an open-weight TTS model with word-level tone control; @bertgodel's team shipped Ultramarine 1 Flash, claiming performance similar to Fable 5 at 60x lower cost after roughly $300k of post-training by five people (weights in two weeks); and @opencode made DeepSeek V4.1 Flash free.
• AGI forecaster @eli_lifland put broadly superhuman AI at ~10-15% by end of 2027 and ~50% by end of 2030, with ~67% AI takeover if labs don't slow down and ~40% extinction conditional on takeover; @Jonas_Vollmer, who calls Eli the forecaster he trusts most, says the post made him more worried.
• @gregisenberg returned from 48 hours in SF with 21 observations, headlined by data brokers reportedly making around $2M a week, agent harnesses as the new GPT wrappers with swappable models, and value migrating from software to the data inside it.
• @ianlivingstone published the first draft of an Agent Baseline Software Factory Reference Architecture, an open blueprint for securing autonomous software factories; @dexhorthy says it turns previously "hand-wavy" MCP gateway talk into a spec builders can actually use.
• A tutorial from @felixrieseberg, an eng lead on Claude, about using AI more ambitiously drew reactions from @tkvtk, @cgtwts, and @thorstenball, who flagged how much first-principles thinking sits inside its simple prompts.

Open Weights Hit Voice, Knowledge Work, and Small GPUs

The day's clearest pattern across the feed: open releases aimed at the niches where paid APIs charge most. @shiri_shh joked that the ElevenLabs CEO opened X to find "his entire product released with open weights," quoting @FishAudio's launch of Drama 3. The announcement, which cites a $52M seed round and 15M+ users, claims "the most emotive and controllable TTS model in the world," open-weight, with control over how a single word sounds, tone changes, and mid-sentence voice shifts.

Knowledge work got similar treatment. @bertgodel announced Ultramarine 1 Flash, a model the team labels "frontier-class" and open-source, claiming "similar performance to Fable 5, while being 60x cheaper," post-trained by five people for ~$300k all-in, available on API now with weights releasing in two weeks. @EXM7777's viral framing goes well beyond the release post: beating OpenAI and Anthropic, under $4M raised, 88 kinds of real tasks from bankers, consultants, and lawyers, and PowerPoint decks, Excel models, and Word memos from a single prompt in the Feather app. Worth noting the gap: the announcement claims parity with Fable 5, not victory over the big labs.

Local options widened too. @opencode announced DeepSeek V4.1 Flash is now free in OpenCode. @coldniko shipped an easy implementation of Qwen3.6-35B-A3B for iGPUs and smaller systems, with the caveat that no big optimizations were done yet, drawing awe from @aktsmm at the sheer development speed. @sudoingX introduced a feed of hands-on local AI experiments across GPU tiers from 8GB up to 256GB, including uncensored and unusual models, aimed at getting "normal people" started with owning their models on their own hardware.

The economics reached game guides as well. @0xFitz open-sourced Polestar on CurseForge: 158 guides covering every race, levels 1-60, "proudly made with AI," after getting fed up with paid competitors Zygor and RestedXP. A small but concrete example of AI-assisted production making a free, open alternative viable against paid products.

Claude's Eng Lead Says Be More Ambitious; Practitioners Answer With Principles

@felixrieseberg, an eng lead working on Claude, published a tutorial on what "being more ambitious" with AI actually means in practice. It spread in three directions: @tkvtk bluntly paraphrased it as an Anthropic engineer saying everyone uses AI badly; @cgtwts shared screenshots of the prompts; and @thorstenball pointed at the deeper lesson, the first-principles reasoning buried in plain instructions, quoting the line "review exists so leadership catches problems early and moves money and people to right places."

@poteto's post on pstack's engineering principles complements the tutorial: very concise skills of roughly 10-20 lines that distill good practice, with a /poteto-mode index that makes agents name "each principle that shaped a decision and the specific choice it changed." The principles came out of months of correcting agents (one, Attack the Premise, was inspired by @VictorTaelin), and @poteto expects most to compress into short principles or generic workflows as models get smarter. @unclebobmartin reckons there's enough wisdom in there to teach a master class.

Isenberg's SF Notes: Value Migrates to Data, Harnesses, and Memory

@gregisenberg's 21-point field report from San Francisco is the densest read of the day. Highlights, all attributed to him: data brokers making around $2M a week selling company data to labs; a dinner question that silenced half a table, "If a new model drops next month, does your company get better or worse?"; a revenue bar of $1B in 12 months paired with almost nobody discussing retention; the mental model of the model as CPU, harness as operating system, skills as apps; a guess that an agent which watches you work for a week and then does your job is about 12 months out; harness teams treating models as swappable and preferring portable prompts over fine-tuning; suites of $5B+ apps cloned open source, pushing value toward the customer data and workflows inside legacy SaaS; internal data cleanup before MCP exposure as a consulting opportunity; agent failure logs as a valuable asset; memory as the next consumer fight, where "the model that remembers the most about you wins"; agent insurance arriving once an agent makes an expensive mistake; and the most-wanted hire being someone with taste who can write evals.

Physical AI threads through the post: @gregisenberg relays a robotics founder's line that "the hardest problems in AI are deployment problems now, not research problems." Separately, @naveedg recommended following @Ryan__Dunwoody, an 11-year Shenzhen hardware veteran ($40M raised, 500K+ custom products shipped, a Guinness record for the first 3D-printed laptop) now running Powerhouse and a hardware accelerator, and who is asking for leads in hardware, physical AI, and robotics.

First Reference Architecture for Securing Autonomous Software Factories

@ianlivingstone announced the first draft of the Agent Baseline Software Factory Reference Architecture, described as "an opinionated blueprint for securing increasingly autonomous software factories," open for feedback and contributions. @dexhorthy supplied the backstory: a podcast conversation with @vtahowe, @ianlivingstone, and the CISO of a billions-of-market-cap company about giving developers MCP access to ticketing, email, and Slack without inviting data exfiltration, prompt injection, or credential hijacks. The previously hand-wavy "MCP gateway" idea, he argues, has now materialized into a formal spec, a starting point rather than a finish line. A small sign the vocabulary is circulating: @jamescmartinez polled software factory people on their favorite software factory.

Also in Tech: Forced PS5 Updates, Browser Piracy, and a Roadster Claim

@Macrike reports a forced update to PS5 firmware 14.10, with PSN access disabled on 14.00 and below. @pokerdev7 argues piracy will reach unprecedented levels as browser play removes the risky-download barrier that used to deter people, quoting @izzynobre's incredulous reaction at whatever browser-play mechanism started the thread. And @elonmusk called the new Tesla Roadster "a candidate for greatest product demo of all time."

Practical Takeaway

If you pay for AI APIs, this week's strongest theme is testability: before renewing spend on voice or knowledge-work endpoints, benchmark the open-weight options that just landed (Drama 3 now, Ultramarine 1 Flash weights when they ship in two weeks) against your current stack, and keep @gregisenberg's dinner question in mind, whether a new, cheaper model makes your product better or worse. If you direct agents daily, borrow from @poteto and @felixrieseberg: capture your hard-won lessons as short first-principles rules, and require the agent to name which rule shaped each decision, so the reasoning surfaces instead of hiding in the output.