Kim 分析近期 AI 前沿动态：GPT-5.6 已完成训练两个月并开放早期访问，延迟发布或因 OpenAI 与政府及监管机构的早期协调。Anthropic 在此方面表现较差。未来模型审查趋严可能导致官方发布进一步延迟。可靠消息源 Andrew Curran 指出，GPT-6 可能在未来六周内发布预览版或正式版，该模型采用全新预训练，发布节奏加速。前沿模型如 Fable 5 和 GPT-5.6 在复杂或智能体任务中 token 消耗显著增加。Sam Altman 称 GPT-5.6 token 效率提升 54%，但总计算需求仍持续增长，推理芯片重要性上升。计算和能源将成为近期最大瓶颈。美国数据中心资本支出今年超 8000 亿美元，中国在能源方面拥有优势，可能限制西方访问其前沿模型。
A few thoughts on the very near future
First of all， what had previously been little more than a rumor has now been confirmed： GPT-5.6 had already been fully trained for two months and was available to selected users in early access. The obvious question is why it was not rolled out earlier.
I do not think this was because OpenAI feared that the model might be overshadowed by Fable 5 or Mythos 5. Instead， OpenAI likely began working with government and regulatory authorities at a very early stage to ensure that the model could be released at all. Even after it had been previewed and announced， it still took some time before it could be rolled out publicly. That said， OpenAI clearly handled the rollout far better than Anthropic， which apparently did not have the same level of cooperation with government and regulatory authorities.
Conversely， however， this also clearly means that future delays and increasingly strict model reviews will probably force us to wait longer for official releases.
The next widely discussed rumor is that， within a few weeks， most likely no more than six， we will see either a preview or even the release of GPT-6. （Andrew Curran @AndrewCurran_ is one of the most reliable sources here on X， so I think that's very realistic.） The model has undergone entirely new pretraining， and the pace of releases is accelerating. The numbers are clear： Frontier labs are releasing more and better models at an increasingly rapid pace. Whereas we once had to wait months， quarters， or even half a year for major new releases， they are now arriving almost weekly.
The latest frontier models may be more efficient in terms of intelligence per token， but they are also being deployed with much larger reasoning budgets. In practice， models such as Fable 5 and GPT-5.6 often consume considerably more tokens during complex or agentic tasks.
Kim 分析近期 AI 前沿动态：GPT-5.6 已完成训练两个月并开放早期访问，延迟发布或因 OpenAI 与政府及监管机构的早期协调。Anthropic 在此方面表现较差。未来模型审查趋严可能导致官方发布进一步延迟。可靠消息源 Andrew Curran 指出，GPT-6 可能在未来六周内发布预览版或正式版，该模型采用全新预训练，发布节奏加速。前沿模型如 Fable 5 和 GPT-5.6 在复杂或智能体任务中 token 消耗显著增加。Sam Altman 称 GPT-5.6 token 效率提升 54%，但总计算需求仍持续增长，推理芯片重要性上升。计算和能源将成为近期最大瓶颈。美国数据中心资本支出今年超 8000 亿美元，中国在能源方面拥有优势，可能限制西方访问其前沿模型。
A few thoughts on the very near future
First of all， what had previously been little more than a rumor has now been confirmed： GPT-5.6 had already been fully trained for two months and was available to selected users in early access. The obvious question is why it was not rolled out earlier.
I do not think this was because OpenAI feared that the model might be overshadowed by Fable 5 or Mythos 5. Instead， OpenAI likely began working with government and regulatory authorities at a very early stage to ensure that the model could be released at all. Even after it had been previewed and announced， it still took some time before it could be rolled out publicly. That said， OpenAI clearly handled the rollout far better than Anthropic， which apparently did not have the same level of cooperation with government and regulatory authorities.
This is not necessarily a sign of declining efficiency. Rather， it suggests that improvements in efficiency are being reinvested into deeper reasoning， longer trajectories and more capable agentic behavior. The result is that total compute consumption per task can continue to rise even as the underlying models become more efficient. Fable 5 and GPT 5.6 demonstrate just how intensive token usage has become. Although Sam Altman explicitly stated that GPT-5.6 is 54% more token-efficient （via CNBC）， the fact remains that compute demand continues to increase， requiring more powerful and efficient computing infrastructure. Inference chips will probably become even more important as well.
In summary， my initial conclusion from the latest releases is that compute demand will not merely continue to grow， but will probably exceed the available supply. This naturally means that energy demand will also increase， and， based on my initial assessment， probably more sharply than previously expected. This is likely to remain the largest bottleneck in the very near future. And this is important to me： there are bottlenecks. Not the training of the models， but besides compute， above all energy. This needs to be taken seriously！
The US power grid， for example， is a major bottleneck， and the obvious question is how the necessary expansion can be achieved. Capital expenditure on data centers in the United States continues to rise sharply. This year， it exceeds 800 billion. It is not yet clear what the situation will look like in 2027， but I can hardly imagine investment declining or less CapEx being required. The reason lies precisely in the developments already mentioned： Demand is growing， particularly demand for energy.
China clearly has an advantage here， a genuine moat， and I believe the West must be extremely careful not to fall behind because of the energy advantage China already possesses in practice. This could also help explain why， according to a recent Reuters report， China is considering restricting Western access to its frontier models. It may have concluded that it will win the long-term race.
Unless there is a genuine breakthrough， whether in small modular nuclear reactors or fusion energy， I expect major problems to emerge over the coming years， for example by 2030. So far， I do not see any viable solutions.
We can therefore clearly establish two points：
Models are becoming larger， better， and increasingly useful for all users. There is no end to this development in sight.
At the same time， the bottleneck appears to be growing increasingly severe， and this is already visible in practice.
Regulation， energy demand， and compute demand could mean that， in the very near future， the release cadence will not accelerate as quickly as hoped or desired. This creates a clear contradiction.
Conversely， however， this also clearly means that future delays and increasingly strict model reviews will probably force us to wait longer for official releases.
The next widely discussed rumor is that， within a few weeks， most likely no more than six， we will see either a preview or even the release of GPT-6. （Andrew Curran @AndrewCurran_ is one of the most reliable sources here on X， so I think that's very realistic.） The model has undergone entirely new pretraining， and the pace of releases is accelerating. The numbers are clear： Frontier labs are releasing more and better models at an increasingly rapid pace. Whereas we once had to wait months， quarters， or even half a year for major new releases， they are now arriving almost weekly.
The latest frontier models may be more efficient in terms of intelligence per token， but they are also being deployed with much larger reasoning budgets. In practice， models such as Fable 5 and GPT-5.6 often consume considerably more tokens during complex or agentic tasks.
This is not necessarily a sign of declining efficiency. Rather， it suggests that improvements in efficiency are being reinvested into deeper reasoning， longer trajectories and more capable agentic behavior. The result is that total compute consumption per task can continue to rise even as the underlying models become more efficient. Fable 5 and GPT 5.6 demonstrate just how intensive token usage has become. Although Sam Altman explicitly stated that GPT-5.6 is 54% more token-efficient （via CNBC）， the fact remains that compute demand continues to increase， requiring more powerful and efficient computing infrastructure. Inference chips will probably become even more important as well.
In summary， my initial conclusion from the latest releases is that compute demand will not merely continue to grow， but will probably exceed the available supply. This naturally means that energy demand will also increase， and， based on my initial assessment， probably more sharply than previously expected. This is likely to remain the largest bottleneck in the very near future. And this is important to me： there are bottlenecks. Not the training of the models， but besides compute， above all energy. This needs to be taken seriously！
The US power grid， for example， is a major bottleneck， and the obvious question is how the necessary expansion can be achieved. Capital expenditure on data centers in the United States continues to rise sharply. This year， it exceeds 800 billion. It is not yet clear what the situation will look like in 2027， but I can hardly imagine investment declining or less CapEx being required. The reason lies precisely in the developments already mentioned： Demand is growing， particularly demand for energy.
China clearly has an advantage here， a genuine moat， and I believe the West must be extremely careful not to fall behind because of the energy advantage China already possesses in practice. This could also help explain why， according to a recent Reuters report， China is considering restricting Western access to its frontier models. It may have concluded that it will win the long-term race.
Unless there is a genuine breakthrough， whether in small modular nuclear reactors or fusion energy， I expect major problems to emerge over the coming years， for example by 2030. So far， I do not see any viable solutions.
We can therefore clearly establish two points：
Models are becoming larger， better， and increasingly useful for all users. There is no end to this development in sight.
At the same time， the bottleneck appears to be growing increasingly severe， and this is already visible in practice.
Regulation， energy demand， and compute demand could mean that， in the very near future， the release cadence will not accelerate as quickly as hoped or desired. This creates a clear contradiction.
Chubby♨️@kimmonismus · X