Really cool! Switched out my default terminal to this and am using it for the last half an hour. Posted a few feedback comments. Looking forward to it developing more :)
Would this still hold true in your opinion if models like O3 become super cheap and bit better over time? I don't know much about the AI space, but as a vanilla backend dev also worry about the future :)
I was helping a relative still in college with a project, and I was struck by how lackadaisical they are about cut-and-pasting huge chunks of code from chatgpt into whatever module they are building without thinking about why, or what it does, or where it fits, as long as it works. It doesn't help that it's all relatively same-looking Javascript so frontend or backend is kinda mixed together. The troubleshooting help I provided was basically untangling the mess by going from first principles and figuring out what goes where. I can tell you I did not feel threatened by the AI there at all, if anything I felt bad for the juniors and feeling like this is what we old people are going to end up having to support very soon.
Not sure how accurate these numbers are but on https://openrouter.ai/ highest used "apps" basically can auto-accept generated code and apply it to the project. I was recently looking at top performers on https://www.swebench.com/ and noticed OpenHands basically does the same thing or similar. I think the trend is going to get much worse, and I don't think Moore's Law is going to save us from the resulting chaos.
We know that OpenAI is verz good at least in one thing: generating hype. When Sora was announced everyone thought that this will be revolutionary. Look at how it looks like in production. Same when they started floating rumours that they have some AGI prototype in their labs.
They are the Tesla of the IT world, overpromise and under deliver.
It's a brilliant marketing model. Humans are inherently highly interested in anything which could be a threat to their well-being. Everything they put out is a tacit promise that the viewer will soon be economically valueless.
I hope people will come to the realisation that we have created a good plagiarizer at best. The "intelligence" originates from the human beings who created the training data for these LLMs. The hype will die when reality hits.
Hype is very interesting. The concept of Hyperstition describes fictions that make themselves real. In this sense, hype is an essential part of capitalism:
"Capitalization is [...] indistinguishable from a commercialization of potentials, through which modern history is slanted (teleoplexically) in the direction of ever greater virtualization, operationalizing science fiction scenarios as integral components of production systems." [0]
"Within capitalist futures markets, the non-actual has effective currency. It is not an "imaginary" but an integral part of the virtual body of capital, an operationalized realization of the future." [1]
This corresponds to the idea that virtual is opposed to actual, not real.
Religion too. Those that are told a prophecy is to come, have a lot of incentive to fulfill that prophecy. Human belief systems are strange and interesting because (IMO) of the entanglement of beliefs with identity
Generally speaking, I think it would. I’m open to being wrong. I think there is a non-trivial amount of hype around O3, and while it would certainly be interesting if it was cheap, I don’t think it would address important issues that AI currently doesn’t seem to even begin to accommodate in its current capacity to recognize or utilize contexts.
For example, I have little to no expectation that it will handle software architecture well. Especially refactoring legacy code, where two enormous contexts need to be held in mind at once.
I'm really curious about something, and would love for an OpenAI subscriber to weigh in here.
What is the jump to O1 like, compared to GPT4/Claude 3.5? I distinctly remember the same (if not even greater) buzz around the announcement of O1, but I don't hear people singing its praises in practice these days.
I gave up interest in GPT4/Claude3.5 about 6 months ago as not very helpful, producing plausible but wrong code.
Have an o3-mini model available to me on the other hand I'm very impressed with its fast, succinct, correct answers while tooling around in zsh on my mac. what things are called, why they exist. why is macports installing db48 etc. It still fails to write simple bash one liners. (I wanted to pipe the output of ffmpeg to a column of --enabled-features and it just couldn't do it)
It's a very helpful rubber duck but still not going to suffice as an agent, but I think its worth a subscription. I wanted to do everything local and self hosted and briefly owned a $3000 mac studio to run llama3.3-70B but it was only as good as GPT4 and too slow to be useful so returned it. In that context even $200/m is relatively cheap.
I don't know how to code in any meaningful way. I work at a company where the bureaucracy is so thick that it is easier to use a web scraper to port a client's website blog than to just move the files over. GPT 4 couldn't write me a working scraper to do what I needed. o1 did it with minimal prodding. It then suggested and wrote me a ffmpeg front-end to handle certain repetitive tasks with client videos, again, with no problem. Gpt4 would often miss the mark and then write bad code when presented with such challenges
He can abstract more models in general frameworks (This is true for every job). He just is hesitant taking the leap to go all Camus and derive money from absurdism ;)
> Tasks that are very context specific or require fine-tuning. For instance, if you're making a RAG system for medical documents, the embedding space is best when it creates larger deviations for the difference between seemingly-related medical words.
Foundational models (GPT-4, Llama 3 etc) effectively compress “some” human knowledge into its neural network weights so that it can generate outputs from inputs.
However, obviously it can’t compress ALL human knowledge for obvious time and cost reasons, but also on the basis that not all knowledge is publicly available (it’s either personal information such as your medical records or otherwise proprietary).
So we build Retrieval Augmented Gen AI, where we retrieve additional knowledge that the model wouldn’t know about to help answer the query.
We found early on the LLMs are very effective at in-context learning (look at 1-shot, few-shot learning) and so if you can include the right reference material and/or private information, the foundational models can demonstrate that they’ve “learnt” something and answer far more effectively.
The challenge is how do you the right content to pass to the foundational model? One very effective way is to use vector search, which basically means:
Pass your query to an embedding model, get a vector back. Then use that vector to perform a cosine-similarity search on all of the private data you have, that you’ve previously also generated an embedding vector for.
The closest vectors are likely to be the most similar (and relevant) if the embedding model is able to generate very different vectors for sources that superficially, seemingly related topics but are actually very very different.
A good embedding model returns very different vectors for “University” and “Universe” but similar for “University” and “College”
Classical word embeddings are static - their value doesn't change depending on the context they appear in.
You can think of the word embedding as a weighted average of embeddings of words which co-occur with the initial word.
So it's a bit of a blurry meaning.
Is "bark" related to a dog? Or to a tree?
Well, a bit of both, really. The embedding doesn't care about the context of the word - once it's been trained.
So if you search for related documents based on word embeddings of your query - it can happen that you miss the mark. The embeddings simply don't encode the semantics you need.
In fact, this can happen even with contextual embeddings, when you look for something specific or in a specialized domain. With word embeddings it's just much more apparent.
For the last ~3 years I worked for companies. No proper publication. But some articles here and there [0]. My Master’s thesis [T]. A (trivial) game to learn Ruby [1]. There are a lot that are not published anywhere.
I have left my full time job this month, and now pursuing full time research independently hoping some papers + job exp. + GRE lands me a PhD in a good research uni.
When you are working, try to make sure that the job aligns with your long term goals. Then learn a lot from work. Learning at work is effortless.
Thank you. That’s very helpful. My day job is pretty basic stuff so I’m trying to learn things on the side, but have the same worry that the day job is where the main learning opportunities would be, so I probably need to find a better gig. Thanks, and good luck on your PhD search.
