LIVE: OpenAI Shutting Down Sora!! It's OVER. Also I'm running Opus 4.6 locally????
TL;DR
OpenAI is killing the Sora app, and Alex Finn thinks it blew its own momentum — he says Sora was genuinely useful for quickly making dumb meme videos of friends, but the moment OpenAI removed direct sharing and forced posting to a feed, he deleted it and stopped caring whether it lived or died.
The real OpenAI rumor is bigger than Sora: a unified app and a new model codenamed “Spud” — Alex says the chatter he’s hearing is that Atlas, Codex, and ChatGPT may get merged, with “Spud” potentially landing in about a month and even being considered as GPT-6 or an AGI-level leap.
He demoed a locally running “distilled Opus 4.6” on a Mac Studio and it looked surprisingly legit — the model is really a Qwen 3.5-based distillation, but when he asked it for a polished Snake game in one HTML file, it produced a glossy, sound-enabled result that felt much closer to Opus than the usual cheap local slop.
Alex’s core bet is that open-source AI agents will beat closed systems for real autonomy — he likes Claude Code and thinks Anthropic has “nailed coding,” but for agents that actually do work on your machine, he says he wants something model-agnostic and customizable like OpenClaw, not a locked-down product such as Perplexity Computer or Claude’s built-ins.
His most actionable advice was to build products for problems AI can’t solve yet — his argument is that if you build the “harness” now for impossible tasks like one-click advanced 3D game creation, then when AGI or a stronger model like Spud arrives, you just swap the API and suddenly your product works before everyone else reacts.
The stream’s funniest mini-drama was Alex unpacking Peter Steinberger publicly correcting him on X — after Alex tweeted that OpenAI “bought OpenClaw,” Peter replied that OpenAI bought neither the project nor the foundation, only hired him; Alex conceded the semantics but argued the tweet was “spiritually” right and wished Peter had just DM’d him instead of dunking in public.
The Breakdown
Sora dies, and Alex isn’t exactly mourning
Alex opens with “breaking news” energy: OpenAI is shutting down Sora, and he immediately frames the real question as whether this is the end of AI slop. But his take is more personal than industry-wide — he loved Sora at launch because it let him generate ridiculous videos of friends and send them around as memes. Once OpenAI removed direct sharing and pushed people toward posting on a feed, he says they “had gold” and threw it away, so he deleted the app and never looked back.
The rumor mill moves to “Spud” and the OpenAI super-app
From there he pivots into what he’s “hearing at the Alex Finn AI news desk”: a new OpenAI model codenamed Spud, potentially dropping in around a month. He says the theory is that Atlas, Codex, and ChatGPT may be merged into one app, with Spud maybe becoming GPT-6 and maybe inching toward AGI territory. He’s skeptical people actually want everything crammed together, though — when he’s coding in Codex, he doesn’t want Sora and browser tools mashed into the same experience.
Running a distilled “Opus 4.6” locally on a Mac Studio
The stream’s hands-on hook is Alex showing a locally running distilled version of Claude Opus 4.6 on his Mac Studio, fueled by coffee and very little sleep. He’s careful to say it’s not literally Opus — it’s a Qwen 3.5-based distillation trained on Opus-style outputs — but he keeps stressing that it “sounds like Opus 4.6.” The payoff comes later when he asks it for a stylish Snake game in one HTML file and gets back a polished result with glow effects and sound, which visibly surprises him.
Smart glasses, rings, and the dream of Tony Stark OpenClaw
One of the most Alex Finn tangents in the stream is his plan to hack Even Realities G2 smart glasses so he can run OpenClaw through them. He doesn’t care about seeing texts, calendars, or teleprompter features on his face; the whole point is pairing the glasses with a phone and using the included ring to issue commands hands-free. His fantasy is pure Tony Stark: walking around, lifting at the gym, and saying “Henry, build this feature” while the agent works in the background.
Why open-source agents win, and why Hermes still annoys him
Alex argues that for autonomous agents, open source is where the real leverage is because it lets you swap models, tools, and workflows however you want. That’s why he’s more excited by OpenClaw than sanitized products like Perplexity Computer, even while praising Claude Code as one of the best coding breakthroughs so far. He also talks through Hermes, saying it has impressive self-improvement and tool-making abilities, but the “memory compactions” are so violent that the agent forgets where it is mid-task — like waking up in Tenet with an oxygen mask on.
Build for impossible problems now, then swap in AGI later
The strongest section of the stream is basically a startup sermon: build products for challenges current models still can’t solve. His example is a one-click tool that creates advanced 3D multiplayer games — impossible right now, but still a product people would obviously want. His point is that most people will wait for AGI, while the smarter move is to build the interface, workflow, and “harness” now so that when a model like Spud or Opus 5 lands, you plug it in and instantly have something powerful.
The Peter Steinberger callout and Alex’s “spiritually correct” defense
Late in the stream he addresses viewers asking about Peter Steinberger correcting him on X. Alex had praised OpenClaw for continuing to support Anthropic, Gemini, and Cursor “despite the fact that OpenAI bought OpenClaw,” and Peter replied that OpenAI did not buy the project; OpenClaw is owned by the independent OpenClaw Foundation he leads. Alex admits Peter is semantically right, but says the substance of his point still stands because OpenAI effectively hired the person most associated with the project, and he thinks a DM or softer public correction would have been more respectful than hitting him with “don’t spread misinformation.”