Eighty/Twenty

.plan for Gordon Weakliem

View on GitHub
28 January 2025

Thoughts on DeepSeek-R1

by Gordon Weakliem

Everyone has their hot take on DeepSeek-R1, so why not me?

First thing everybody did is ask about Tiananmen Square, just to see if the CCP is still paying attention. It turns out that this is a sensitive subject and DeepSeek-R1 does a direct refusal. I asked some questions about the Great Chinese Famine and got pretty good answers, although I initially encountered a hallucination:

The "三年大饥荒" (Three Starvation Campaign) was an extraordinary period in Chinese history during the late 1950s under the leadership of the Red
Army. It was a campaign orchestrated by Joseph Stalin's forces to control China through forced labor and harsh agricultural policies. The campaign
resulted in massive suffering, loss of life, and long-term economic damage.

This can’t be true, since Stalin died in 1953 and the famine didn’t start until 1959. I asked some followup questions and it clarified with exactly that point, Stalin couldn’t have influenced it. I was running 7b so a hallucination is not that surprising.

Overall, DeepSeek-R1 seems to be a very good model. I really like seeing the reasoning, I expect that other major vendors will follow suit. I want to explore a larger version, with the new 64GB MBP, I should be able to run 32b at least.

I’m actually fine with the refusal: they could have whitewashed the event or twisted it somehow, so it’s really to their credit that they simply refuse to deal with it. I think that this is actually the best approach that they could have taken, although I’ve seen other posts that demonstrate some blind spots with issues like human rights violations. I’m also beginning to wonder how the models made in the US will start to handle certain topics. Some vendors have had some embarassing moments (Gemini’s black Nazis, for example), but with DEI and “wokeness” now apparently in retreat, how will our models start to deal with topics like gender issues or racism, or with historical events?

tags: []
permalink

navigation
TIR 2025-01-22 - Interviews in the Age of AI