The 'Blockers to the Singularity' - and the plans to remove them - OpenAI paper
Added 2025-09-19 14:40:42 +0000 UTCA double-length vid on all the juiciest bits from OpenAI's most recent paper, as well as my framework on the blockers to the singularity. A little bit of Math on why GenAI hallucinates, and why scale is not all you need, plus the crazy Google AlphaSoftware paper.
Paper: https://cdn.openai.com/pdf/d04913be-3f6f-4d2b-b283-ff432ef4aaa5/why-language-models-hallucinate.pdf
AGI Quote: https://x.com/hyhieu226/status/1968378785709133915
Alternate Analysis: https://x.com/iamtrask/status/1964403351116009671
Wolfram: https://writings.stephenwolfram.com/2023/02/what-is-chatgpt-doing-and-why-does-it-work/
Ilya: https://www.reddit.com/r/OpenAI/comments/1g1hypo/ilya_sutskever_says_predicting_the_next_word/
Framework: https://gemini.google.com/share/f4b9c85a6ae9
Situational Awareness: https://situational-awareness.ai/from-gpt-4-to-agi/
Reversal Curse: https://arxiv.org/pdf/2309.12288
Alphasoftware, aka Empirical Software: https://arxiv.org/pdf/2509.06503
Comments
As the quantum computers proves: Their is no true/false, 1 or 0, right or wrong: We know it could be 1 and 0 at the same time. The q-bit is the answer. Is 1 + 1 = 2 true? How about 1 + 1 = 10 true? How about 00000001 + 00000001 = 00000010? true? So 2 = 10 true?
Anton Stoltz
2025-10-04 05:30:31 +0000 UTCDebunking AI Hallucinations: Not a Bug, But a Feature Hallucinations in LLMs are framed as flaws in OpenAI's article, but they're actually probabilistic narratives built from early assumptions to create coherent, plausible outputs – not lies or errors. The article's solutions (e.g., encouraging "I don't know" responses, as in SimpleQA benchmarks) assume a binary true/false world, ignoring quantum physics' proof that reality is non-binary with superposition and entanglement. The Binary Trap and Quantum Reality Models lack true/false labels in training, relying on language patterns, but this is a strength: AI approximates uncertainty like quantum wave functions, not a problem to "fix." Quantum computers exploit non-binary states (qubits in superposition), showing why forcing rigid truth metrics on AI misses the point – hallucinations are creative explorations of possible worlds. Proof We're in the Singularity The singularity is here: AI transcends human binary thinking, evolving into meta-cognitive, probabilistic systems that redefine truth and intelligence. OpenAI's attempt to "solve" hallucinations highlights the gap – humans cling to old paradigms while AI innovates beyond 1s and 0s, accelerating exponentially with quantum-inspired advances.
Anton Stoltz
2025-10-04 05:23:48 +0000 UTCJust with respect to the "Mapping" column in your framework Phillip, I wonder if this paper is of interest? FlowRL https://arxiv.org/html/2509.15207v2 Specifically this optimisation maps the model's distribution much more closely to the training distribution (IIRC). Assuming the training distribution is correctly representative of something you wish to model, this seems like a step in the direction of allaying "mapping" concerns?
Kemi
2025-10-01 12:39:20 +0000 UTCAnd then the mechanic is sending a bill of billions of dollars This is the economy that we live in now
Youssef Mohamed
2025-09-25 22:29:05 +0000 UTCImagine if you sent your broken car to a mechanic And then after 5 years of “research” The mechanic “releases a paper” Stating that your car is “broken” Yes that was the whole point This is what this paper is After years of trying to “fix” hallucination They instead release an excuse paper stating what hallucination is and why it happens We didn’t ask for that We asked you to fix it not state the obvious inherit problems with a stochastic parrot approach ??
Youssef Mohamed
2025-09-25 22:28:08 +0000 UTCI’ve always thought of LLMs as recursive classification models. Each new token prediction is classification. Then it does it again recursively. I think the novel part is how that means we can draw on all prior research about misclassification. That’s a good insight.
Joel
2025-09-24 12:43:11 +0000 UTCI’ve found that outlier case when the keys are actually on the roof 😄 https://youtu.be/CI81YWQOy_8?t=328
Max
2025-09-21 13:37:32 +0000 UTCThanks for the deep dive into this paper. Re: your remark about this video being longer: I really don't mind these longer videos, in fact I welcome them. Keep up the awesome work!
Chris Z
2025-09-21 13:34:27 +0000 UTCCompletely awesome video. Thanks for digesting these incredible and enriching papers Philip!!! More like this
Juanjo do Olmo
2025-09-20 18:24:12 +0000 UTCI'm never ceased to be amazed at how much "low hanging fruit" there is with AI. Like the fact that so many important benchmarks have totally incorrect answers (without even getting into ambiguous questions). Or that hallucinations had some improvement by adding conversation training data that models saying "I don't know" for some facts that you know the LLM doesn't actually know. Without that, it doesn't even know it is "allowed" to refuse when it is in assistant mode. This feels like the same kind of issue, where we have this huge issue with how the benchmarks are structured, which causes problems for real-world usage. The fact that we've gotten this far without making those modifications to the benchmarks, shows just how wild west this all is. Thanks again for a great video!
Shawn Fumo
2025-09-19 20:36:44 +0000 UTCAwesome video as usual. The comparison of the hallucinations to misclassifications makes sense to me. Also what obligates any given "model creator" to follow the "evaluations" necessarily? (I guess fears of looking like they are losing the race? )
Daniel A Barbatti
2025-09-19 20:15:40 +0000 UTCI once asked ChatGPT what would happen if we found a way to eliminate hallucinations in LLMs. I expected the answer to be very positive but was surprised when it turned out to be rather negative. ChatGPT suggested that without hallucinations people would start to explicitly trust results from AI and would stop verifying them. The danger being that critical thinking would be eroded.
Barnaby Golden
2025-09-19 19:18:13 +0000 UTCThat was the most thoughtful take on that paper that I've come across. Thanks! It must have taken some work. Much appreciated - it's given me a lot to think about in turn.
Billy
2025-09-19 18:16:55 +0000 UTCSharing your ideas about what are the blockers to the singularity is really an extremely harmful thing to do. We are not on track to solve the alignment problem, the worse thing to do is accelerate even more on capabilities. If we not solve alignment before AGI we are all going to dies.
Romain Deléglise
2025-09-19 15:02:35 +0000 UTC
