Sam Altman's Final Battle: Delivering AGI in 4 Months - Exclusive Insight - 36氪

Sam Altman's Final Battle: Deliver AGI in 4 Months 新智元 2026-08-30 09:27 The AGI competition is heating up, tech giants are sprinting to secure key positions, and the winner could take all. Some people have already announced that AGI has been achieved. Just at the earnings call this Wednesday, Jensen Huang dropped a remark: For many tasks, we can say that we have already achieved AGI. On the same day, a more than two-hour long interview with Sam Altman was also released. He set the timeline at the end of 2026. For Sam Altman, these four months are far more than a technical sprint. Before the company goes public next year, he must first play the AGI card. If he fails to deliver, by the time he rings the listing bell, there will already be a rival that has gone public earlier, boasts twice the market value, and is already profitable standing in front of him. In this battle, OpenAI has no way to retreat. Jensen Huang spoke first: AGI has been achieved At that earnings call, someone asked him what he thought of companies like OpenAI chasing AGI so desperately. Jensen Huang's answer was that there is nothing to argue about on this issue for now. The entire industry has not even reached a consensus on what "intelligence" is, let alone what standard to use to measure AGI. He holds something else in his hand. NVIDIA's AVO architecture achieved a full 100% score on ARC-AGI-3 , a test specifically designed to assess whether agents can work autonomously for long periods of time. Following this result, he put forward an extremely ambitious prediction: In the near future, NVIDIA may only need to maintain around 40,000 human employees, but they will have 400,000 or even 4 million digital employees. But that full-score test paper is nothing but a superficial shell As is known to all, the gameplay of ARC-AGI-3 is quite "unconventional". Agents are directly thrown into an unfamiliar game, with no rules or targets provided at all. They have to explore, observe and guess completely on their own. They only have a 64×64 grid and several buttons. Each environment has at least six levels: the first level can be passed in five steps, while the sixth level requires 50 steps to complete. It is difficult even for humans to play, and models are not much better. The key point here: the model used by AVO is Claude Opus 5. Then NVIDIA added a shell to it. Not a single weight parameter was modified, and the score directly jumped from 30.16% to full marks. Two things actually play a role here. The first is persistent memory. The memory will not be cleared even when the context is full, and the agent continues to work from the current state without stepping into the same pit again. The second is the supervisor, which does not do the work itself, but only monitors the entire search trajectory. Once it finds the agent is going in circles, it will pull the main agent to other strategies. However, strictly speaking, this full-score result is not actually valid. More than three hours after NVIDIA's official announcement, François Chollet, the father of ARC-AGI, posted on Twitter that: Getting 100% on the public demo set does not mean getting 100% on the ARC-AGI-3 benchmark. It is like saying you have cleared a video game, but in fact you only finished the tutorial level. According to the ARC Prize's own description, the public set is not designed for scoring at all. The semi-private set and private set are the only valid parts, and AVO's transcript has no scores for these two parts. What is more subtle is that this is already the third full-score result in six weeks. Tycho took the lead at the end of July, VISTA got another full score on August 5, and AVO is the third. The three parties have completely different solutions, but all the driving models are Claude Opus 5. More and more people are announcing that AGI has been achieved, but it seems that only those two companies are still capable of actually building AGI. And no one has actually played this card yet. Why does Sam Altman dare to say the end of the year? In the exclusive interview, Sam Altman was full of confidence: By the end of the year, there will be a system inside OpenAI that he is willing to call AGI. Chief Research Officer Mark Chen is relatively more conservative. In his view, OpenAI is still 20% away from AGI. This statement sounds arrogant, but they do have something in their hands. This year, OpenAI set a very clear internal goal: to create an automated AI research intern that can actually do real work. And this "AI automated intern" is Astra. Give Astra an experimental idea, and it can write code, run experiments, and submit results on its own in OpenAI's code repository. Throw a research paper to it, and Astra can complete the workload that a top human researcher takes about a week to finish. 16 agents held a roundtable meeting At an internal demonstration not long ago, dozens of executives of major customers witnessed this scene with their own eyes. On the big screen, 16 AI agents split a research-level math problem into a bunch of sub-problems. Each of them took their own task, cross-checked with each other halfway, and finally pieced together a complete proof. It looks exactly like a team with tacit cooperation holding a closed-door meeting. The second demonstration is even more shocking. Astra opened common desktop software, jumped back and forth between different applications, created files, modified content, and scheduled tasks so fast that it did not look like a human typing on the keyboard at all. All the people who saw it exclaimed that Astra has fully realized the full capability of "persistent agents" this time. Sam Altman even predicted: "I expect this will be the first model to truly invent important new things. This is a very AGI-like thing." Pay attention to the wording "invent new things", which means creating something completely unknown to all mankind out of thin air. 9 months to build a chip manually, Astra personally took part in the work Recently, OpenAI's first self-developed chip "Jalapeño" directly outperformed Jensen Huang's flagship GPU by a large margin. The number one hero behind this is GPT-Astra. It got involved directly, working all the way from circuit design to software development. On the chip side, it optimized the underlying circuit. On the software side, it wrote and optimized the most underlying core Kernel, and also tuned three major open source models by the way. As a result, the performance of the BF16 multiplier surged by 56%, while the area of the matrix unit was reduced by 10%. The two core modules of "attention mechanism" and "MoE" are 1.5 to 1.8 times faster than the versions handwritten by human experts. Ladies and gentlemen, this is AI modifying circuits at the physical level. Interestingly, NVIDIA is doing exactly the same thing. AVO is not built for games at all. Its main battlefield is writing low-level code for GPUs. In March this year, NVIDIA published a paper, whose core idea is to hand over the entire "code modification" step to an agent that can work autonomously. In the actual test, it ran autonomously for 7 consecutive days without any human intervention, explored more than 500 optimization directions, and finally submitted 40 valid kernel versions. The generated core code is up to 3.5% faster than NVIDIA's own closed-source version, and up to 10.5% faster than the industry's most cutting-edge open source implementation. AI is building AI Now, connecting these two things together, Sam Altman's words make sense. Every model iteration consumes a huge amount of human research time. But if Astra starts taking over this part, the entire R&D speed will change qualitatively. Once this cycle starts running, the only thing left is a race for speed. The final battle has fully begun In fact, GPT-5.6 and Fable 5 that we can use publicly are all products of the last round. And according to leaks, the real next-generation models of both sides have actually finished training. OpenAI: Behind Astra, there is another model called Bel in the pipeline On OpenAI's side, according to leaks from Leo and others who have long tracked the dynamics of cutting-edge models, the pre-training arrangements this year are very dense: Spud (Sol): Pre-training completed in March, corresponding to the currently public GPT-5.5 and 5.6 series; Doug: Completed from April to May. Based on it, post-training such as reinforcement learning was carried out to produce the upcoming Astra, which is known as GPT-6 to the outside world; Bel: Pre-training just ended in August, with a total parameter count of more than 10 trillion, and a scale close to that of GPT-4.5 back then. The leak says that OpenAI internally regards Bel as "the foundation model after GPT-6", and it may even become the base of a model that crosses the AGI threshold. Anthropic: Fable 5.1 is already in gray release On Anthropic's side, someone found on August 18 that some accounts on the Claude web version that selected Fable 5 were quietly switched to a new model. This is Anthropic's old rule before release, and Fable 5 itself was rolled out in this way in the past. 该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。