OpenAI's 2023 Leadership Crisis, Q* Speculation and Synthetic Data

Several stories are colliding: Sam Altman's return to OpenAI, reports about a project called Q*, Elon Musk's comments on copyrighted training data, and the potential of synthetic data. I think they may be connected, but the connections I propose are hypotheses rather than facts established by the available public evidence.

OpenAI's November 29, 2023 announcement says Altman is returning as CEO, Mira Murati as CTO and Greg Brockman as president. The initial board consists of Bret Taylor as chair, Larry Summers and Adam D'Angelo. The company says it will work on governance and an independent review of recent events ([OpenAI's announcement](https://openai.com/index/sam-altman-returns-as-ceo-openai-has-a-new-initial-board/)).

Altman says Ilya Sutskever will no longer serve on the board and that they hope to continue working together. I speculate that a major AI discovery and disagreement over whether to commercialize it triggered the firing. The announcement does not say that, and Sutskever's departure from the board did not itself establish that he was leaving the company.

I also wonder whether Musk might try to recruit Sutskever, given their prior relationship. That is my conjecture, not evidence of a job offer. I read Altman's language as an attempt to repair a difficult relationship, but the personal motives behind the board's actions are not established by this note.

Altman thanks the outgoing directors and Emmett Shear, who briefly served as interim CEO, and emphasizes his partnership with Brockman. I discuss D'Angelo's continued board role and his connection to Poe. Those observations do not by themselves establish a conflict of interest or explain the board decision.

The leadership list in the announcement does not include Sutskever, which I interpret as a possible sign of a future departure. Altman also praises the team and says no employee was lost during the episode. My interpretation of the missing name is not a confirmed employment decision.

Altman lists research and safety, product deployment, and rebuilding the board and governance among the immediate priorities. I infer that safety may have driven the dispute, but the statement itself does not support that conclusion. In March 2024, OpenAI's summary of an independent review said the prior board's decision did not arise from product safety, security or development pace concerns ([review summary](https://openai.com/index/review-completed-altman-brockman-to-continue-to-lead-openai/)).

In a November 29 interview with The Verge, Altman describes being asked to return and says he initially needed a few minutes to work through the emotions. When pressed about why he was fired, he does not give a detailed explanation and welcomes an independent investigation. I remain curious and speculate about a Q* connection, but the interview does not verify one.

Asked about reports of a Q* breakthrough, Altman says he has no particular comment on the 'unfortunate leak' and speaks generally about rapid progress and safety. I take the wording as acknowledgment of a leak, but he does not confirm the reported system's architecture, capabilities or role in the governance crisis. The broader claims in the video's title should be read with that limit ([The Verge interview](https://www.theverge.com/2023/11/29/23982046/sam-altman-interview-openai-ceo-rehired)).

In an excerpt from Andrew Ross Sorkin's interview, Elon Musk alleges that AI companies train on copyrighted data and disputes contrary claims. That is Musk's allegation. Whether particular uses infringe copyright or qualify for a legal exception depends on facts and law and cannot be settled by the interview alone.

Sorkin asks about articles and quotations uploaded to X and what their aggregation could mean for original rights holders. I argue that X gives Musk access to a valuable stream of user-created material and suggest that restricting API access reflects its strategic value. Those are my interpretations; individual quotations being posted on a platform does not automatically resolve the copyright status of their use for training.

Musk says he expects AI progress to outpace the resolution of lawsuits and describes becoming fatalistic about AI risk. The interview also displays his prediction that AI could surpass the smartest human within three years. That was a forecast made in 2023, not an established timetable or a reason to assume any legal dispute has become irrelevant.

I connect Musk's forecast to my theory that OpenAI could use synthetic data to advance beyond models trained on existing material. I think this might be related to Q*, but neither the interview nor the public material shown demonstrates that connection. The claim that legal outcomes will no longer matter is likewise a prediction, not a legal conclusion.

Nathan Lambert's [essay on synthetic data](https://www.interconnects.ai/p/llm-synthetic-data) provides the next piece of the discussion. Synthetic data is produced by a machine rather than directly authored by a human; generated instructions, preferences and critiques can supply examples for training or evaluation. It is not inherently accurate simply because it can be produced at scale.

I argue that human-created text is finite and that generated material could greatly expand the volume available for training. I speculate that Q* may generate high-quality data and use it for self-improvement, but this is a hypothesis without confirmation in the sources presented. More data alone does not guarantee better models or AGI.

Lambert discusses the prospect of using AI-generated feedback in alignment work, including OpenAI's then-new Superalignment effort. I see the appeal of reducing a human feedback bottleneck, while worrying about handing over too much oversight. Automating feedback creates its own questions about what the evaluator rewards and whether mistakes or biases are amplified.

One argument for synthetic data is that more diverse examples could improve performance on rare tasks. A counterargument is that outputs generated by existing models may carry their limitations and biases rather than automatically push the state of the art forward. I lean toward the concern that generated examples can be derivative, while acknowledging that the outcome depends on how they are generated, filtered and evaluated.

I compare the issue with AlphaGo learning from games against itself. Go offers a clear win-or-loss signal, making iterative evaluation more direct than for open-ended language. The historical account needs a qualification: the original AlphaGo used both human game data and reinforcement learning through self-play before its famous victory over Lee Sedol; AlphaGo Zero later trained without human game examples ([Google DeepMind](https://deepmind.google/research/alphago/)).

With language models, it can be much harder to decide which answer is better. Automated benchmarks help, but contamination occurs when evaluation examples or closely related material appear in training data, making a score less informative about generalization. My own test prompts are one way to inspect behavior, though a small personal test is not immune to limitations either.

I return to open models: groups without a proprietary data source or large budget may find generated data especially attractive. Lambert notes examples such as Alpaca and Vicuna that used synthetic instruction data to fine-tune Llama-based models. These show a use of generated examples, not proof that synthetic data alone can build any frontier model.

I plan to explore synthetic data and benchmark contamination in more detail. The possibility of rapid progress is exciting to me and also raises concerns about safety and oversight. My optimism about the future is a view, not evidence that a particular forecast or rumored breakthrough has been realized.