171–180 of 268 posts
I agree with Simon’s article but I usually think about “research” to mean comparing different kinds of evidence (not just the search part). Like evidence for the effectiveness of Obamacare. Or how some legal case may play out in the courts. Or how much The Critic influenced The Family Guy. Or even what the best way to use X feature of Y library. I’ve found ChatGPT and other LLMS can struggle to evaluate evidence - to…
I guess the part where I'm still skeptical are: Google is also still pretty good at search (especially if I avoid the AI summary with udm=14). I'll take one of your examples: Britannica to seed Wikipedia. I searched for "wikipedia encyclopedia brtannica". In less than 1 second, I got search results back. I spend maybe 30 seconds scanning the page; past the Wikipedia article on Encyclopedia Britannica, past the Encycl…
I guess the part where I'm still skeptical are: Google is also still pretty good at search (especially if I avoid the AI summary with udm=14). I'll take one of your examples: Britannica to seed Wikipedia. I searched for "wikipedia encyclopedia brtannica". In less than 1 second, I got search results back. I spend maybe 30 seconds scanning the page; past the Wikipedia article on Encyclopedia Britannica, past the Encycl…
First, you not having to spend the 60 seconds and it means you can parallelize it with something else to get the answer effectively instantly. Second, you're essentially establishing that if an LLM can get it done in less than 60 seconds its better than your manual approach, which is a huge win, as this will get faster!
> Starbucks in the UK don’t sell cake pops! Do a deep investigative dive ... I used to play games on my computer a lot. Not so much anymore, don't really want to lock myself in a room alone and play games. I have kids and a wife, and it feels isolative. But those days I would, and often the hardware I had was underpowered to be able to experience the game in its full glory. I would often spend hours and hours just ho…
This is a common theme with LLMs (and LLM criticism). The context: I was rushing for a train, I ran into Starbucks at the station for a coffee, I noticed they didn't have cake pops and the staff member didn't appear to know what they were. I see three choices here: 1. Since I'm mildly curious about Starbucks and cake pop availability in the UK, I get on the train, open up my laptop and dedicate realistically a solid…
Alternatively, you could have spent that half hour on the train exercising your own creativity to try and satisfy your curiosity. Whether you're right or wrong doesn't really matter, because as you acknowledge it's not really important enough to you to matter. Picking (2) eliminates all the possible avenues that might have lead you down.
I'm not saying one is better than the other, just that you're approaching the criticism on the basis of axioms that represent a narrow viewpoint: That of someone who has to be "right" about the things they are curious about, no matter how trivial.
There is one big failure mode though - ChatGPT hallucinates middle of simple textual OCR tasks!
I will feed ChatGPT a simple computer hardware invoice with 10 items - out comes perfect first few items, then likely but fake middle items (like MSI 4060 16GB instead of Asus 5060 Ti 16GB) and last few items are again correct.
If you start prompting with hints, the model will keep making up other models and manufacturers, it will apologize and come up with incorrect Gigabyte 5070.
I can forgive mistaking 5060 for 5080 - see https://www.theguardian.com/books/booksblog/2014/may/01/scan... . However how can the model completely misread the manufacturers??
This would be trivially fixed by reverting to Tesseract based models like ChatGPT used to do.
PS Just tried it again and 3rd item instead of correct GSKILL it gave Kingston as manufacturer for RAM.
Basically ChatGPT sort of OCRs like a human would, by scanning first then sort of confabulating middle and then getting the footer correct.
I agree with Simon’s article but I usually think about “research” to mean comparing different kinds of evidence (not just the search part). Like evidence for the effectiveness of Obamacare. Or how some legal case may play out in the courts. Or how much The Critic influenced The Family Guy. Or even what the best way to use X feature of Y library. I’ve found ChatGPT and other LLMS can struggle to evaluate evidence - to…
If you ask for information which is e.g. academic or technical it would cite information and compare different results, etc, without any extra prompt or reminder.
Grok 4 (at the initial release) was just reporting information in the articles it found without any analysis.
Claude Opus 4 also seems bad: I asked it to give a list of JS libraries of a certain kind in deep research mode, and it returned a document focused on market share and usage statistics. Looks like it stumbled upon some articles of that kind and got carried away by it. Quite bizarre.
So GPT-5 is really good in comparison. Maybe not perfect in all situations, but perhaps better than an average human
I do miss the earlier "heavy" models that had encyclopedic knowledge vs the new "lighter" models that rely on web search. Relying on web search surfaces a shallow layer of knowledge (thanks to SEO and all the other challenges of ranking web results) vs having ingested / memorized basically the entirety of human written knowledge beyond what's typically reachable within the first 10 results of a web search (eg: digiti…
From what I can tell, they are pretty damn big.
Grok 4 is quite large too.
I guess the part where I'm still skeptical are: Google is also still pretty good at search (especially if I avoid the AI summary with udm=14). I'll take one of your examples: Britannica to seed Wikipedia. I searched for "wikipedia encyclopedia brtannica". In less than 1 second, I got search results back. I spend maybe 30 seconds scanning the page; past the Wikipedia article on Encyclopedia Britannica, past the Encycl…
I find ChatGPT to be great at research too-but there are pathological failure modes where it is biased to shallow answers that are subtly wrong, even when definitive primary sources are readily available online: https://www.fortressofdoors.com/researchers-beware-of-chatgp...
There is value in pruning the search tree because the deeper nodes are usually not reputable. I know you have cause to believe that "Wilhelm Matzat" is reputable but I don't think it can be assumed generally. If you were to force GPT to blindly accept counter points from people - the debate would never end. And there has to be a pruning point at which GPT would accept this tradeoff: maybe the less reputable or well known sources may have a correct point at the cost of being incorrect more often due to taking an incorrect analysis from a not well known source.
You could go infinitely deep into any analysis and you will always have seemingly correct points on both sides. I think it is valid for GPT to prune the search at a point where it converges to what society at large believes. I'm okay with this tradeoff.
I do miss the earlier "heavy" models that had encyclopedic knowledge vs the new "lighter" models that rely on web search. Relying on web search surfaces a shallow layer of knowledge (thanks to SEO and all the other challenges of ranking web results) vs having ingested / memorized basically the entirety of human written knowledge beyond what's typically reachable within the first 10 results of a web search (eg: digiti…
I keep switching between both but I think I'm starting to prefer the lighter one that is based on the sources instead.