My take is that Altman recognizes LLM winter is coming and is trying to entrench.
They just won’t be the hottest thing since smartphones.
"write a TCL/tk script file that is a "frontend" to the ls command: It should provide checkboxes and dropdowns for the different options available in bash ls and a button "RUN" to run the configured ls command. The output of the ls command should be displayed in a Text box inside the interface. The script must be runnable using tclsh"
It didn't get it right the first time (for some reason wants to put a `mainloop` instruction) but after several corrections I got an ugly but pretty functional UI.
Imagine a Linux Distro that uses some kind of LLM generated interfaces to make its power more accessible. Maybe even "self healing".
LLMs don't stop amazing me personally.
Current LLMs being 80% to being 100% useful doesn't mean there's only 20% effort left.
It means we got the lowest-hanging 80% of utility.
Bridging that last 20% is going to take a ton of work. Indeed, maybe 4x the effort that getting this far required.
And people also overestimate the utility of a solution that's randomly wrong. It's exceedingly difficult to build reliable systems when you're stacking a 5% wrong solution on another 5% wrong solution on another 5% wrong solution...
>And people also overestimate the utility of a solution that's randomly wrong. It's exceedingly difficult to build reliable systems when you're stacking a 5% wrong solution on another 5% wrong solution on another 5% wrong solution...
I call this the merry go round of hell mixed with a cruel hall of mirrors. LLM spits out a solution with some errors, you tell it to fix the errors, it produces other errors or totally forgets important context from one prompt ago. You then fix those issues, it then introduces other issues or messes up the original fix. Rinse and repeat. God help you if you don't actually know what you are doing, you'll be trapped in that hall of mirrors for all of eternity slowly losing your sanity.
I wrote some data visualizations with Claude and aider.
For anything that someone would actually pay for (expecting the robustness of paid-for software) I don’t think we’re there.
The devil is in the details, after all. And detail is what you lose when running reality through a statistical model.
It’s not like it’ll do it consistently.
Just a marketing stunt.
It was trained to generate text in the lean language (https://www.lean-lang.org/) which is specifically used for formal proofs.
It’s not a natural language model.
Source: https://deepmind.google/discover/blog/ai-solves-imo-problems...
Google seems to mainly be playing the game of more specialized models (AlphaGo, AlphaProof) with general training methods (AlphaZero)
I do think it’s kind of funny that they mention AGI in that article, but the model is specifically not general.
"oi i need lik a scrip or somfing 2 take pic of me screen evry sec for min, mac"
with an actual (and usually functional) script to be "glorified grammar corrector", then sure.
I think actually the best use case for LLMs is "explainer".
When combined with RAG, it's fantastic at taking a complex corpus of information and distilling it down into more digestible summaries.
I think that RAG and RAG-based tooling around LLMs is gonna be the clear way forward for most companies with a properly constructed knowledge base but I wonder what you mean by "explainer"?.
Are you talking about asking an LLM something like "in which way did the teams working on project X deal with Y problem?" and then having it breaking it down for you? Or is there something more to it?
1. I got this medical provider that has a webapp that downloads graphql data(basically json) to the frontend and shows some of the data to the template as a result while hiding the rest. Furthermore, I see that they hide even more info after I pay the bill. I download all the data, combine it with other historical data that I have downloaded and dumped it into the LLM. It spits out interesting insights about my health history, ways in which I have been unusually charged by my insurance, and the speed at which the company operates based on all the historical data showing time between appointment and the bill adjusted for the time of year. It then formats everything into an open format that is easy for me to self host. (HTML + JS tables). Its a tiny way to wrestle back control from the company until they wise up.
2. Companies are increasingly allowing customers to receive a "backup" of all the data they have on them(Thanks EU and California). For example Burger King/Wendys allow this. What do they give you when you request data? A zip file filled with just a bunch of crud from their internal system. No worries: Dump it into the LLM and it tells you everything that the company knows about you in an easy to understand format (Bullet points in this case). You know when the company managed to track you, how much they "remember", how much money they got out of you, your behaviors, etc.
2. I think if the data was significantly large, the llm would alias a ton of potentially important info.
I don't understand enough about #2 to comment, but it's certainly interesting.
Some trials have their protocols published.
Here's an example trial: https://clinicaltrials.gov/study/NCT06613256
And here's the protocol: https://cdn.clinicaltrials.gov/large-docs/56/NCT06613256/Pro... It's actually relatively short at 33 pages. Some larger trials (especially oncology trials) can have protocols that are 200 pages long.
One of the big challenges with clinical trials is making this information more accessible to both patients (for informed consent) and the trial site staff (to avoid making mistakes, helping answer patient questions, even asking the right questions when negotiating the contract with a sponsor).
The gist of it here is exactly like you said: RAG to pull back the relevant chunks of a complex document like this and then LLM to explain and summarize the information in those chunks that makes it easier to digest. That response can be tuned to the level of the reader by adding simple phrases like "explain it to me at a high school level".
The last system, I led one team competing for the Transcelerate Shared Investigator Portal (we were one of the finalist vendors).
Little side project: https://zeeq.ai
I just tried Gemini and it was useless.
Someone says "X is the model that really impressive. Y is good too."
Then someone responds "What?! I just used Z and it was terrible!"
I see this at least once in practically every AI thread
Their laughably overzealous nanny-state censorship, paired with a model so appallingly inept it would embarrass a chatbot from the 90s, makes it nothing short of highway robbery that this digital dumpster fire is permitted to masquerade as a product fit for public consumption.
The sheer gall of Google to foist this steaming pile of silicon refuse onto unsuspecting users borders on fraudulent.