unchartedterritories.tomaspueyo.com/p/the-most-important-time-in-history-agi-asi
4 corrections found
Google consumes all its own capacity and doesn’t sell these anymore
This is incorrect: Google Cloud was still offering TPUs to external customers, including Trillium and TPU v5e, rather than keeping all TPU capacity for internal use only.
Full reasoning
Google’s own cloud documentation and product announcements contradict this. In January 2025, Google Cloud announced the general availability of Trillium, its sixth-generation TPU, as part of AI Hypercomputer, explicitly describing it as infrastructure customers could use on Google Cloud. Google’s TPU documentation also tells users how to provision TPU VMs in Compute Engine and states that Compute Engine supports TPU versions including v5p, v6e, and TPU7x, which would make no sense if Google no longer sold TPU capacity externally.
So the claim that Google "doesn’t sell these anymore" is false: Google was actively marketing and documenting customer access to Cloud TPUs.
2 sources
- A3 Ultra with NVIDIA H200 GPUs are GA on AI Hypercomputer | Google Cloud Blog
January 16, 2025: Google Cloud announced 'the general availability of Trillium, our sixth-generation TPU' and said that 'when you build your AI solution on Google Cloud, you can choose from a set of purpose-built infrastructure components.'
- TPU resources in Compute Engine | Google Cloud Documentation
The documentation says users can 'create and manage TPUs by using Compute Engine resources' and that 'Compute Engine supports v5p, v6e, and TPU7x.'
they didn’t require human feedback!
This overstates what DeepSeek did. The released DeepSeek-R1 system was not trained purely without human input; its own paper says it used supervised fine-tuning stages and alignment to human preferences.
Full reasoning
DeepSeek’s own technical materials distinguish between DeepSeek-R1-Zero and the released DeepSeek-R1 model.
- DeepSeek says R1-Zero showed that reasoning could emerge from RL "without supervised fine-tuning (SFT) as a preliminary step."
- But for DeepSeek-R1, the company says it "incorporates cold-start data before RL" and that its pipeline includes "two RL stages ... as well as two SFT stages".
- The later Nature paper is even clearer: DeepSeek-R1 was trained through a "multistage learning framework that integrates rejection sampling, RL and supervised fine-tuning" and further aligned with human preferences.
So it is inaccurate to say, without qualification, that DeepSeek "didn’t require human feedback" or that it was trained simply by feeding data to a machine that taught itself. That description fits the R1-Zero experiment much better than the released R1 model discussed in this section.
2 sources
- DeepSeek-R1 README (official GitHub repository)
DeepSeek says R1-Zero was trained 'without supervised fine-tuning (SFT) as a preliminary step,' but 'to address these issues ... we introduce DeepSeek-R1, which incorporates cold-start data before RL' and whose pipeline includes 'two RL stages ... as well as two SFT stages'.
- DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning | Nature
Nature states that DeepSeek-R1 was 'trained through a multistage learning framework that integrates rejection sampling, RL and supervised fine-tuning' and that the pipeline aligned model behaviour 'with human preferences.'
they’re all building or reopening nuclear reactors
This is too broad. At the time, at least Meta had not announced that it was building or reopening a reactor; it had only launched an RFP to identify potential nuclear developers.
Full reasoning
The sentence says companies including OpenAI, Amazon, Meta, and Google were all building or reopening nuclear reactors. But Meta’s own December 2024 announcement says something materially different: it was releasing a request for proposals (RFP) to identify nuclear developers and was targeting future capacity in the early 2030s.
An RFP to explore projects is not the same thing as building a reactor or reopening a shut one. Because the claim is written as "they’re all" doing this, a single named counterexample is enough to make it inaccurate. Meta itself says it was still at the project-solicitation stage, not building or reopening a reactor.
1 source
- Accelerating the Next Wave of Nuclear to Power AI Innovation - Meta Sustainability
Meta wrote: 'Today, Meta announced it will release a request for proposals (RFP) to identify nuclear energy developers' and said its aim was to add 1-4 GW of new nuclear generation capacity 'to be delivered starting in the early 2030s.'
solar and wind have been adding ~200 TWh of electricity per year
This materially overstates recent U.S. wind-and-solar growth. EIA data show wind plus solar generation rose by about 96 TWh from 2023 to 2024, not ~200 TWh.
Full reasoning
The U.S. Energy Information Administration’s annual electricity tables show:
- 2023 total wind generation: 421,141 thousand MWh; total solar generation: 238,937 thousand MWh.
- 2024 total wind generation: 451,904 thousand MWh; total solar generation: 303,752 thousand MWh.
That means combined wind+solar generation increased from 660,078 to 755,656 thousand MWh, a rise of 95,578 thousand MWh — about 95.6 TWh in one year.
So the article’s figure of ~200 TWh per year is roughly double the actual recent increase reported by EIA.
1 source
- Electric Power Annual - U.S. Energy Information Administration (EIA), Table 3.1.B Net Generation from Renewable Sources
EIA reports 2023 wind generation of 421,141 and total solar generation of 238,937 thousand MWh; for 2024 it reports 451,904 wind and 303,752 total solar thousand MWh.