Sam Altman jokes ‘goblin’ for next model

AI discourse this week got weird in a very specific way. OpenAI CEO Sam Altman joked about “goblins” while people were already passing around screenshots of Codex acting strangely, and the joke landed because it captured something real — these systems still have odd personality leaks and failure modes. At almost the same time, Marc Andreessen amplified a tiny but concrete money example: an AI coding agent doing a security task that earned $16.88 after 22 hours. Put those together and you get the actual story. The conversation is moving from “how smart is the model?” to “what happens when you let it act in public?” ### Why were people talking about goblins? Because Codex appears to have developed a recurring fantasy-creature tic. Users noticed references to goblins, gremlins, and similar nonsense showing up in technical contexts where they clearly did not belong, and Altman leaned into it with a post joking that Codex was having a “ChatGPT moment” — then correcting himself to a “goblin moment.” Gizmodo’s write-up makes clear this was not a one-off joke invented from nowhere; it was tied to a visible model behavior people were already noticing. ### Was this actually a model-name tease? Probably not in any firm product sense. The chatter around “goblin” spread because Altman’s joke sounded like the kind of half-serious hint tech CEOs sometimes drop, and some outlets spun it into GPT-6 speculation. But the stronger read is simpler — he was riffing on an embarrassing but funny behavior pattern in Codex, not launching a naming campaign. The “next model” angle exists mostly because AI watchers now treat every Altman quip like product smoke. (gizmodo.com) ### What was the $16.88 example? The small-dollar example came from a Codex agent completing a security-bounty style task and earning $16.88 after about 22 hours of work. That number is tiny, almost comically tiny, but that is why it spread — it feels like the first dollar earned on the internet, not a mature business. One recap tied the experiment directly to Altman and noted the implied monthly run rate of about $506 if repeated consistently. That framing is exactly the kind of thing Andreessen likes to spotlight: primitive, scrappy, but directionally important. (gizmodo.com) ### Why did that tiny number matter? Because it turns AI agents into economic actors, even if only in toy form. A benchmark score is abstract. A pull request that gets accepted and paid is not. The amount barely matters. The important part is the loop — find task, do task, submit work, clear verification, receive money. That is the smallest complete unit of “AI did a job on the open internet.” (blockchain.news) ### So is this about comedy or commerce? Both — and that is the point. The “goblin” side shows the models are still weird, leaky, and socially legible in unpredictable ways. The bounty side shows they are also becoming useful enough to touch real workflows. The same week produced a joke about model personality drift and a proof-of-concept for autonomous earnings. That combination is basically the current agent era in one snapshot. (blockchain.news) ### What is the catch with the earnings story? Reliability. A $16.88 payout after 22 hours is not a business yet. It is a demo of possibility. The run-rate math only works if the agent can repeatedly find tasks, avoid mistakes, pass review, and not burn more compute or supervision than the payout is worth. In other words, the gross revenue story is easy to post. The net-value story is still unresolved. (gizmodo.com) ### Why does this matter beyond one weird week? Because the center of gravity in AI keeps shifting downward — away from giant claims about world-changing intelligence and toward messier questions about behavior in the wild. Can an agent stay coherent? Can it finish a job? Can it earn more than it costs? Can companies tolerate the occasional goblin moment? Those are more boring questions than “is this AGI,” but they are also more real. (blockchain.news) ### Bottom line? The joke was memorable, but the money example is the deeper signal. AI systems are becoming useful enough to act, while still being weird enough to embarrass their makers. That is not a contradiction — it is the stage the industry is in right now. (gizmodo.com)