The Nonlinear Library LW - AI #50: The Most Dangerous Thing by Zvi
Feb 8, 2024
36:57
Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: AI #50: The Most Dangerous Thing, published by Zvi on February 8, 2024 on LessWrong.
In a week with two podcasts I covered extensively, I was happy that there was little other news.
That is, until right before press time, when Google rebranded Bard to Gemini, released an app for that, and offered a premium subscription ($20/month) for Gemini Ultra.
Gemini Ultra is Here
I have had the honor and opportunity to check out Gemini Advanced before its release.
The base model seems to be better than GPT-4. It seems excellent for code, for explanations and answering questions about facts or how things work, for generic displays of intelligence, for telling you how to do something. Hitting the Google icon to have it look for sources is great.
In general, if you want to be a power user, if you want to push the envelope in various ways, Gemini is not going to make it easy on you. However, if you want to be a normal user, doing the baseline things that I or others most often find most useful, and you are fine with what Google 'wants' you to be doing? Then it seems great.
The biggest issue is that Gemini can be conservative with its refusals. It is graceful, but it will still often not give you what you wanted. There is a habit of telling you how to do something, when you wanted Gemini to go ahead and do it. Trying to get an estimation or probability of any kind can be extremely difficult, and that is a large chunk of what I often want. If the model is not sure, it will say it is not sure and good luck getting it to guess, even when it knows far more than you.
This is the 'doctor, is this a 1%, 10%, 50%, 90% or 99% chance?' situation, where they say 'it could be cancer' and they won't give you anything beyond that. I've learned to ask such questions elsewhere.
There are also various features in ChatGPT, like GPTs and custom instructions and playground settings, that are absent. Here I do not know what Google will decide to do.
I expect this to continue to be the balance. Gemini likely remains relatively locked down and harder to customize or push the envelope with, but very good at normal cases, at least until OpenAI releases GPT-5, then who knows.
There are various other features where there is room for improvement. Knowledge of the present I found impossible to predict, sometimes it knew things and it was great, other times it did not. The Gemini Extensions are great when they work and it would be great to get more of them, but are finicky and made several mistakes, and we only get these five for now. The image generation is limited to 512512 (and is unaware that it has this restriction).
There are situations in which your clear intent is 'please do or figure out X for me' and instead it tells you how to do or figure out X yourself. There are a bunch of query types that could use more hard-coding (or fine-tuning) to get them right, given how often I assume they will come up. And so on.
While there is still lots of room for improvement and the restrictions can frustrate, Gemini Advanced has become my default LLM to use over ChatGPT for most queries. I plan on subscribing to both Gemini and ChatGPT. I am not sure which I would pick if I had to choose.
Table of Contents
Don't miss the Dwarkesh Patel interview with Tyler Cowen. You may or may not wish to miss the debate between Based Beff Jezos and Connor Leahy.
Introduction. Gemini Ultra is here.
Table of Contents.
Language Models Offer Mundane Utility. Read ancient scrolls, play blitz chess.
Language Models Don't Offer Mundane Utility. Keeping track of who died? Hard.
GPT-4 Real This Time. The bias happens during fine-tuning. Are agents coming?
Fun With Image Generation. Edit images directly in Copilot.
Deepfaketown and Botpocalypse Soon. $25 million payday, threats to democracy.
They Took Our Jobs. Journalists and lawyers.
Get In...
In a week with two podcasts I covered extensively, I was happy that there was little other news.
That is, until right before press time, when Google rebranded Bard to Gemini, released an app for that, and offered a premium subscription ($20/month) for Gemini Ultra.
Gemini Ultra is Here
I have had the honor and opportunity to check out Gemini Advanced before its release.
The base model seems to be better than GPT-4. It seems excellent for code, for explanations and answering questions about facts or how things work, for generic displays of intelligence, for telling you how to do something. Hitting the Google icon to have it look for sources is great.
In general, if you want to be a power user, if you want to push the envelope in various ways, Gemini is not going to make it easy on you. However, if you want to be a normal user, doing the baseline things that I or others most often find most useful, and you are fine with what Google 'wants' you to be doing? Then it seems great.
The biggest issue is that Gemini can be conservative with its refusals. It is graceful, but it will still often not give you what you wanted. There is a habit of telling you how to do something, when you wanted Gemini to go ahead and do it. Trying to get an estimation or probability of any kind can be extremely difficult, and that is a large chunk of what I often want. If the model is not sure, it will say it is not sure and good luck getting it to guess, even when it knows far more than you.
This is the 'doctor, is this a 1%, 10%, 50%, 90% or 99% chance?' situation, where they say 'it could be cancer' and they won't give you anything beyond that. I've learned to ask such questions elsewhere.
There are also various features in ChatGPT, like GPTs and custom instructions and playground settings, that are absent. Here I do not know what Google will decide to do.
I expect this to continue to be the balance. Gemini likely remains relatively locked down and harder to customize or push the envelope with, but very good at normal cases, at least until OpenAI releases GPT-5, then who knows.
There are various other features where there is room for improvement. Knowledge of the present I found impossible to predict, sometimes it knew things and it was great, other times it did not. The Gemini Extensions are great when they work and it would be great to get more of them, but are finicky and made several mistakes, and we only get these five for now. The image generation is limited to 512512 (and is unaware that it has this restriction).
There are situations in which your clear intent is 'please do or figure out X for me' and instead it tells you how to do or figure out X yourself. There are a bunch of query types that could use more hard-coding (or fine-tuning) to get them right, given how often I assume they will come up. And so on.
While there is still lots of room for improvement and the restrictions can frustrate, Gemini Advanced has become my default LLM to use over ChatGPT for most queries. I plan on subscribing to both Gemini and ChatGPT. I am not sure which I would pick if I had to choose.
Table of Contents
Don't miss the Dwarkesh Patel interview with Tyler Cowen. You may or may not wish to miss the debate between Based Beff Jezos and Connor Leahy.
Introduction. Gemini Ultra is here.
Table of Contents.
Language Models Offer Mundane Utility. Read ancient scrolls, play blitz chess.
Language Models Don't Offer Mundane Utility. Keeping track of who died? Hard.
GPT-4 Real This Time. The bias happens during fine-tuning. Are agents coming?
Fun With Image Generation. Edit images directly in Copilot.
Deepfaketown and Botpocalypse Soon. $25 million payday, threats to democracy.
They Took Our Jobs. Journalists and lawyers.
Get In...