Publicly, @scaling01 is a deliberately mythic AI-evals persona. The indexed X profile shows the account joined on Aug. 18, 2024, links the same legal-notice website you gave, and shows roughly 19,000 posts, about 990 following, and about 38.6k followers. The GitHub side runs the bit even harder: display name “Lisan al Gaib,” username “voice-from-the-outer-world,” location “Arrakis,” and public-facing profile/highlights text that includes “lead them to paradise.” My read is simple: benchmark prophet cosplay with real technical content under the robe. (X (formerly Twitter))
The growth was fast enough to make normal career-building look like a quaint village craft. In one post he called a thread “The Lisan al Gaib Timeline” and said it showed how he got about 8,500 followers; in another he said he was 33 followers away from 20,000 and had been at zero nine months earlier. That suggests the core arc pretty clearly: he became a rapid-response interpreter for frontier-model launches, benchmark drama, and the general delirium of AI release season. (X (formerly Twitter))
The public record starts technical early. By Dec. 2024 he was posting detailed notes on the DeepSeek-V3 technical report, focusing on training cost, MoE architecture, context expansion, and post-training. By Jan. 2026 he was posting sweeping forecasts about coding-and-math AGI, multi-agent systems, benchmark saturation, and lab trajectories, while also attacking GPT-5.2 as too slow, too expensive, and too manager-shaped. So this is not just another launch-day screenshot account. He tries to turn hype into mechanisms, token budgets, and failure modes. (Thread Reader App)
That same habit shows up in the eval fights. In June 2025 he replicated the Tower of Hanoi setup from the “Illusion of Thinking” debate and argued that the apparent collapse was partly an artifact of output-length limits and benchmark design. Outside observers later cited his threads as part of the criticism of that paper. More recently, coverage of ARC-AGI-3 described him as a critic arguing that the methodology was designed to produce low scores. So the public persona is not just hype. It is adversarial QA for benchmark claims, delivered with the bedside manner of a brick. (Thread Reader App)
The builder side is real too. The GitHub account has three public repos: a fork of AidanBench, a fork of ARC-AGI-3-benchmarking, and LisanBench. LisanBench is described as a cheap, scalable benchmark based on open-ended word-ladder chains that stress planning, vocabulary depth, memory, attention, constraint adherence, and long-context “stamina.” The README says it cost under about $50 to evaluate dozens of models and says the concept emerged from a human-AI collaboration, with Claude 4 Opus generating candidate ideas and the creator narrowing and implementing the final version. Even the usage terms insist on crediting @scaling01 and explicitly say they are not a standard open-source license, which tells you this is an auteur benchmark, not a sterile committee appliance. (GitHub)
So who is “Lisan al Gaib” in practice? Not a saint, not a neutral analyst, and definitely not a calm institutional voice. More like an unusually high-throughput public evaluator: part benchmark builder, part paper reader, part model-launch sportscaster, part doomer-hype translator. The Dune branding is silly on purpose. Underneath it is a serious habit of asking, “What exactly is being measured here, and does the graph deserve the headline?” (GitHub)
First, because they build. A lot of internet AI criticism is just decorative sneering in front of someone else’s graph. LisanBench is at least an attempt to convert taste into a runnable evaluation artifact, with methodology, code, constraints, and visible assumptions. (GitHub)
Second, because they pressure-test benchmark narratives in public. The Alignment Forum and AI-evaluation roundups referenced his criticism of reasoning-benchmark claims, and ARC-AGI-3 coverage cited him as part of the measurement dispute. That means he is not just reacting to the discourse; he is helping shape the counter-argument when official benchmark stories look too neat. (Alignment Forum)
Third, because the takes travel. VentureBeat described him as a “developer and AI evaluator,” and newsletter or aggregator ecosystems like Techmeme and Latent Space-style daily news repeatedly surface his posts when new models, pricing, or benchmark results land. That is a decent sign he functions as a signal node, not merely as another angry avatar in the swamp. (Venturebeat)
Fourth, because he self-audits. He publicly reviewed his own 2025 predictions, marked the “lab will declare AGI” call as basically a miss, and later wrote “Lisan failed” about his ARC-AGI-2 forecast. That does not make him infallible. It makes him unusually legible, which online is practically a miracle. (X (formerly Twitter))
This is partial, not exhaustive. Many Jan. 2026 predictions are still live as of March 2026, and X is a terrible archive because apparently human civilization decided its public square should also be an obstacle course.
Not a true corpus-level perplexity calculation, because the available archive is messy and incomplete. But qualitatively, these are the highest-z-score bits of public wording I could verify: