Astra, agents, and the PSP model

00:00:04 Friday morning, right before OpenAI's GPT-6 Astra went live, something odd happened. ChatGPT, Claude, Grok, and Cursor all went down at roughly the same time. The Occam's razor explanation is that Azure had an outage — which they did report around the same window.

00:00:22 But then a slightly more entertaining theory started circulating: maybe Astra's first act as a public model was to take out its competitors on launch day. Once the lights came back on, the launch itself was somehow even messier. OpenAI put up their announcement page, embargoed stories dropped from Reuters, CNBC, and The Verge, and then for reasons nobody really explained, they took it down again.

00:00:50 Tech influencers started flexing about how long they'd secretly had access to Astra before any of us did. About ninety minutes later the page came back up, but the model wasn't available to anyone publicly yet. Plus and Pro rollout would happen in the coming days.

00:01:09 By evening, Sam Altman had posted an apology for the messy rollout. When a Pro subscriber asked if they should stay up waiting for it, his reply was something we've all heard after making a desperate late-night plea: go to bed. He also told CNBC that Astra went through a formal review with the Trump administration before release — so apparently even the government got to try AGI before we did.

00:01:37 What is Astra? The official demo explained it clearly enough. Charlie Guo, OpenAI's Developer Experience engineer, walked us through the features. It's billed as the best model in the world for tasks where raw intelligence matters. The headline features are improved computer use and two new API capabilities.

00:01:59 Computer use takes screenshots to track what's happening in an app while keeping that app open in the background. Asynchronous tool calling lets the model keep working on other parts of a task while one call runs. Steering lets you give the model new context or change its direction mid-response without canceling what it's already started.

00:02:23 You hand Astra a photo, tell it to use Krita to paint it in Van Gogh's style with the Golden Gate Bridge in the background, and it navigates the app interface using screenshots while you work on something else. The same ability can fill out API key forms or do QA on a mobile app without your hands on the keyboard.

00:02:45 In terms of numbers: Astra was pre-trained on more than 100,000 GPUs at OpenAI's Stargate site in Texas. It's apparently the first OpenAI model where previous versions didn't do a significant chunk of the supervising during training — meaning they stopped using their older models to help teach it and went full frontier.

00:03:08 That's a meaningful infrastructure shift, even if you can't see it directly. Benchmark results are mixed in ways that matter more than headlines would suggest. OSWorld scores sit at 73% after forty minutes per task, compared to GPT-5.6 Soul's 65% at seventy-five minutes.

00:03:27 Faster and more accurate on desktop work. Arc AGI 3 hit 99%, designed to test generalization rather than memorization. OpenAI's Trust Me Bro benchmarks gave it 100% on exploit bench and 65% on terminal bench science. The one that caught my attention: Astra is reportedly the first model to hit what OpenAI calls the critical cyber threshold in its preparedness framework.

00:03:53 In plain English, it can find and exploit zero-day vulnerabilities autonomously, without a human telling it what to look for. That's not a benchmark anyone wants to hear about their own infrastructure. Pricing sits at $10 per million input tokens and $50 out — same as Anthropic's Fable 5.1 from Tuesday.

00:04:15 The Independent Intelligence Index from Artificial Analysis put Astra at 61, which is exactly the same as GPT-5.6 Soul and five points behind Fable 5.1 at 66. So we're in territory where the difference between what OpenAI calls its best model and what their own previous generation did comes down to single digits on a scaled index.

00:04:38 Meanwhile, the rest of the week kept moving. Anthropic released Fable and Mythos 5.1 on Tuesday — same model under two names, with Mythos refusing to answer anthrax questions. Case studies included a hedge fund called Millennium using it to trace a five-year-old crash bug through compiled vendor code, protein design hitting fifty percent success rate instead of the usual ten, and training a network on thirty-year-old NASA radar data for a new Venus elevation map.

00:05:11 Meta dropped MuSpark 1.3 on Wednesday, their fourth Superintelligence Labs release in five months. Pricing sits at $1.25 input and $4.25 output, with a contributor tier charging ten cents to let them train on whatever you send. The local model pass on all this is straightforward but worth stating plainly: Astra costs real money to run.

00:05:35 Ten dollars per million input tokens adds up fast if you're using it for sustained work — which is exactly why DHH's tweet from Saturday hits a nerve.

00:05:46 DHH posted Saturday: "This is so cool. Local agents, preconfigured. I know @0xsero is also cooking here. We're going to end up with something amazing out of the box very soon!" DHH tracked the convergence between local model inference and agentic tooling. As models get better at understanding codebases, navigating file systems, and chaining API calls, we stop asking whether a local model can do useful work and start asking what infrastructure you need to make that work reliable.

00:06:24 The reference to @0xsero matters because Sero has been building toward this for a while — preconfigured agent tooling that runs locally rather than shipping your data through someone else's API. That's the trade-off: you lose some of the raw model intelligence that Astra or Fable bring to the table, but you gain privacy, cost predictability, and the ability to run things on hardware you control.

00:06:52 Speaking of controlled hardware, someone in the LocalLLaMA subreddit posted Saturday about running a ninety-million-parameter conversational model on a Sony PSP. The PSP launched in 2004. Someone actually figured out how to get a language model onto a twenty-two-year-old handheld gaming device.

00:07:13 The results are exactly what you'd expect from the math: about half a token per second, which means one to three minutes for a reply. The model can generate poems, short stories, and non-functional code. Sometimes it gets things right — like identifying that Apple makes MacBooks — and other times it hallucinates with complete confidence.

00:07:36 It's not useful in any real metric. But the fact that someone built it at all shows where this space sits. The model weights themselves are tiny by modern standards — ninety million parameters compared to Astra's probably hundreds of billions — and running them on 2004 hardware requires everything from custom compilers to memory management tricks that would've been cutting-edge for the era.

00:08:04 The GitHub repo is at thatblend/LLMPSP. It demonstrates how far we've come by showing how close we can still get to working code on dead hardware, and what makes it linger is the contrast with Astra. One system cost more than a hundred thousand GPUs and took weeks to train.

00:08:23 The other fits in a pocket device that originally sold for two hundred and ninety-nine dollars. Both are running conversational language models. Neither one is close to making the other obsolete. There's also this counterintuitive angle: the slower inference on the PSP isn't just a limitation — it's almost a feature for certain kinds of agentic work.

00:08:47 When an agent has minutes rather than seconds to reason through a problem, it can do more internal verification before outputting. The trade-off between speed and thoroughness shows up in ways benchmarks rarely capture. The gitignore-everything-by-default thread from Hacker News landed at about the same time — arguing for flipping the entire .gitignore philosophy.

00:09:13 Alex Pliutau's piece proposes ignoring everything by default and whitelisting only the files you want tracked, using git's built-in negation syntax with the ! operator. The argument is practical: modern projects generate so much local junk — agentic documentation, IDE state, build artifacts — that a whitelist approach feels easier than maintaining an ever-growing deny list.

00:09:39 He references TypeScript-Go's two-hundred-and-seven-line .gitignore as evidence of how bloated these files have become. I can believe that. My own projects hover around eighty lines and that includes only the essentials. This is a small piece of builder infrastructure, not a model story, but it points at the same shift.

00:10:01 As agentic tooling generates more local artifacts — context files, planning documents, scratchpads — we need better defaults for what gets tracked and what stays ephemeral. The tools are evolving faster than the conventions around them.

00:10:17 Robin Hanson posted Saturday what feels like the kind of question he could only ask: "What percent of those who vote in the 2028 US presidential election will ask a large language model for a recommendation?" Two-part question, naturally, because Hanson being Hanson means he follows up with: "What percent of those who ask an LLM how to vote in the 2028 US presidential election will NOT do what the LLM advises?"

00:10:53 His framing treats voter behavior as an information problem — people have preferences but uncertain information about candidates, and they're increasingly willing to outsource the information-gathering step to an AI model. The follow-up question is the sharper one: compliance rate.

00:11:12 Asking for advice is cheap. Following it costs something — cognitive dissonance if the recommendation contradicts your existing leanings, or a genuine preference reversal if it doesn't. We don't have good data on either number yet because voting advice from language models is still nascent, but Hanson's two-part structure reveals the real uncertainty.

00:11:35 Anatoly Karlin replied: "I think the AI will be smarter than me by 2028. I will not only ask but implement as it instructs." He doesn't say he'll consider the advice or factor it into his decision-making. He says he'll act on it. That's a stronger claim than either of Hanson's questions, and it suggests Karlin sees this as a trust transfer rather than an information aid.

00:12:06 The four-year gap matters here too. Hanson is asking about 2028 from a vantage point that gives model developers exactly the kind of runway that makes confident predictions possible — or impossible, depending on how you look at it. If we get to the election with models that are significantly more capable than whatever Astra or Fable represent today, then Karlin's confidence in deferring to them isn't unreasonable.

00:12:34 What I'd actually want to know is whether there's already any empirical signal. Are voters using AI tools for political research today? What fraction of people who ask a large language model a political question actually read the full answer versus just skimming the recommended candidate name?

00:12:54 Those data points don't exist publicly yet, but they're the kind you'd expect to surface within twelve to eighteen months. The other number that matters is adoption among younger voters. Demographic projections show the 2028 electorate skewing younger. A cohort that treats AI-assisted research as normal rather than novel would shift the election's information ecology in ways that veterans of the past three decades haven't modeled yet.

00:13:23 The Hanson-Karlin exchange is worth keeping because it's one of the few places where we can see an economist and a political commentator both landing on something close to agreement about the direction of travel, even if they're measuring different things along it.

00:13:41 Hanson asks what percent will ask and what percent won't comply. Karlin says he'll do both. The real gap between them isn't really substantive — it's that Hanson is asking the question for a population while Karlin is answering for himself. Self-selection bias applies to anyone who voluntarily responds to a tweet about AI voting advice.

00:14:03 Still, four years from now will either confirm or invalidate both of their intuitions. And unlike most political predictions, we have a mechanism to test them: just show up at the polling place on election day and see whether people's behavior matches what they said they'd do.

00:14:22 What ties this week together isn't the release cadence across OpenAI, Anthropic, and Meta. It's the infrastructure question moving to the edge even as the frontier accelerates. Astra cost over a hundred thousand GPUs and took weeks to train. The PSP model fits in a pocket device from 2004 and runs on whatever memory you can scavenge.

00:14:45 Both are running conversational language models, both are doing real work for someone, and neither one is close to making the other obsolete. The gitignore thread at the end of this episode is small — builder infrastructure, not a model story — but it points at the same shift.

00:15:05 As agentic tooling generates more local artifacts (context files, planning documents, scratchpads), we need better defaults for what gets tracked and what stays ephemeral. The tools are evolving faster than the conventions around them. On Astra itself: the 73% on OSWorld in forty minutes versus Soul's 65% at seventy-five is real progress on desktop automation.

00:15:31 The AI Intelligence Index score of 61 matching GPT-5.6 Soul exactly matters — it means the raw intelligence gap between generations is narrowing to within measurement noise, which shifts the competitive advantage from model capability to everything around it: tooling, pricing, reliability, and the kind of async orchestration Astra introduced via its steering and parallel tool calling.

00:15:58 What would change my read on Astra: whether OpenAI follows up with public benchmark results that separate computer-use performance from pure reasoning ability. Right now we have desktop automation scores but no clean separation between how much of that 73% on OSWorld comes from the model being smarter versus the computer-use pipeline being more mature.

00:16:23 That matters because computer use is a feature layer — it's the implementation around the model, not the model itself. On Hanson's question: I'd look for the first poll-based data on how many voters currently consult AI tools during their research phase. If that number jumps from single digits to double digits before 2028, then Karlin's compliance prediction becomes less speculative and more of a self-reported commitment.

00:16:53 Either way, four years is a long runway in this space. We'll see which direction wins out by then. — Seln Oriax.