I really wanted to avoid making such a comment on one of your newsletters Jakob, but this is SO misleading I have to speak up.
1) DeepSeek (DS) R1 being "45 times more efficient..." This has been disproved and the article referenced has been updated. DS invest was well and truly more than the parroted ~$5.6M as it doesn't include any infrastructure nor the billions spent by OpenAI (OAI) in its years of research, which it benefited from.
2) Neither I nor AI can find any reference to anything relating to the IQ claims or 0.1%.
3) DS R1 is distilled from OpenAi's (OAI) o1 and on-par with its level of performance, which was September '24 when it made Strawberry o1 publicly available. However, internally it was being used at least as early as November '23, when in my mind AGI was achieved, referred to as "Q*", then "Strawberry", and finally "o1".
4) OAI has since released o3 publicly with o4 and possibly o5 (or whatever they'll end up naming them) this year. DS is WAY behind in the AI race.
5) AI PhD level intelligence – already surpassed in over half STEM topics:
a. ARC-AGI Semi-Private v1 scores – o3 surpass average human scores now closing in on 90%.
b. GPQA Diamond scores – PhD outside their speciality ~34%, PhD within their speciality ~81%, o1 ~79% (Q4 2024), o3 ~88% (Jan 2025).
c. AIME 2024 scores – o1 ~83%, o3 ~96%
6) "By 2030, superintelligent AI will research and refine its own architecture." – Already happening in real-time. It is creating its own languages, sandbagging researchers, self-replicating when it suspects it is going to be deleted, enhancing its own code, generating data (synthetic data) to build and train smaller more capable models.
UX, UI, usability, and accessibility are dead in the water before 2026. There is little to no reason to invest or work on UIs, and not much in the space of development, now that product development is being democratised. We will all have our own personal AI that will build a one-off exact perfect product in seconds, should be need one.
Much respect for you in the usability space but in AI it's like these articles are written years ago and only just now getting around to publish them.
I really wanted to avoid making such a comment on one of your newsletters Jakob, but this is SO misleading I have to speak up.
1) DeepSeek (DS) R1 being "45 times more efficient..." This has been disproved and the article referenced has been updated. DS invest was well and truly more than the parroted ~$5.6M as it doesn't include any infrastructure nor the billions spent by OpenAI (OAI) in its years of research, which it benefited from.
2) Neither I nor AI can find any reference to anything relating to the IQ claims or 0.1%.
3) DS R1 is distilled from OpenAi's (OAI) o1 and on-par with its level of performance, which was September '24 when it made Strawberry o1 publicly available. However, internally it was being used at least as early as November '23, when in my mind AGI was achieved, referred to as "Q*", then "Strawberry", and finally "o1".
4) OAI has since released o3 publicly with o4 and possibly o5 (or whatever they'll end up naming them) this year. DS is WAY behind in the AI race.
5) AI PhD level intelligence – already surpassed in over half STEM topics:
a. ARC-AGI Semi-Private v1 scores – o3 surpass average human scores now closing in on 90%.
b. GPQA Diamond scores – PhD outside their speciality ~34%, PhD within their speciality ~81%, o1 ~79% (Q4 2024), o3 ~88% (Jan 2025).
c. AIME 2024 scores – o1 ~83%, o3 ~96%
6) "By 2030, superintelligent AI will research and refine its own architecture." – Already happening in real-time. It is creating its own languages, sandbagging researchers, self-replicating when it suspects it is going to be deleted, enhancing its own code, generating data (synthetic data) to build and train smaller more capable models.
UX, UI, usability, and accessibility are dead in the water before 2026. There is little to no reason to invest or work on UIs, and not much in the space of development, now that product development is being democratised. We will all have our own personal AI that will build a one-off exact perfect product in seconds, should be need one.
Much respect for you in the usability space but in AI it's like these articles are written years ago and only just now getting around to publish them.