davidajackson

Does anyone here have a lot of knowledge on neuromorphic computing, or has worked in the space? What do you think of it? Also, do you see any ethical issues with this type of engineering...

In this thread I would like to discuss the interesting ways we've found to increase model accuracy when analyzing images. I have been learning about vision models and trying a few different strategies, to varying degrees of success, and would like to brainstorm a bit further.

To that end, I have a few questions for HN:

What's your experience with bounding boxes?

What's your experience with different models/what are your favorites?

What's your experience with triaging extraction errors while managing cost?

Wanted to have a brief discussion here as to whether you see A2A as the 'final state' for most AI automation. Are there any alternative patterns that you feel might be competitive, or better than A2A? Thoughts on A2E?

Do self driving cars ever make assumptions about the lack of desirability of a lane if there is no one in it? While counterintuitive, sometimes that is an indication that there might be a parked car not currently visible etc. This has struck me as a curious problem where cars would actually observe other car's decisions ahead to guess that 'something is up'.

I have had a theory for a while that economic growth or the economy in terms of CAGR will go to 0% or becomes linear with population as the population grows. This in my mind aligns with the opportunities for lifestyle improvements going to 0 (we already have the internet, we already have food, shelter etc.) so this is no new market to cover/no new businesses to invent?

Has this topic been discussed or debated anywhere? Has anyone with this same theory done any mathematical projections?

Is there evidence that understanding causality (action -> reaction, cause -> effect, etc.) is instinctive or does it need to be trained in humans? It seems obvious that humans need to be taught some forms of logic, like logical fallacies, but are there some that are instinctive? How has this been 'proved'?

Siri is a really bad product when compared to GPT. Why has it taken Apple so long to get AI right?

I see some different names thrown around for this concept. Names aside, will this be the future, or do we anticipate humans sticking with the current hardware path? What are the performance or space tradeoffs of the two different system approaches?

It's been entertaining reading about the OpenAI-Meta talent wars. I thought it would be an interesting discussion to ask some OpenAI employees here how they got an 'in' to one of the most challenging companies to join.

Obviously, with the levels of talent on here, many of us are more founder-oriented/already have great gigs, but I'm curious if there were any special processes or connections you had to get the job. No agenda here :) just want to hear more about this.

Any more details on uses here? I read a post from Stripe on this. Any other companies or any of you using these successfully?

FYI... has anyone here considered how to represent background information (background context) at the beginning of a transformer sequence? Specialized tokens that refer to user attributes (a token preamble, if you will...).

Micropayments? Uselessness? I'd like to hear some thoughts.

Stripe extended migration dates to payment intents, however they killed the ach credit transfer option for sources on August 15 even if extensions were granted. Pretty surprising and would have expected a warning in the extension UI.

Has this decision been thought through a bit more? I know @pc sometimes is on here. Why not extend that endpoint support until after migration deadlines?

I thought I would start a thread on here to discuss some of the challenges in Generative AI particularly with enforcing on-topic conversation and tone. It seems this poses substantial cost restrictions on deployment, as using on topic checks and classifiers for inference prompts does not scale well. It seems an O(N) problem that has high cost and low performance with inference. In addition, though embedding methods could be used to ensure on topic conversations, I want to specifically discuss the problem of combined context, where a predefined cluster is isolated, but combined alternative context in the prompt 'dilutes' the vector such that it no longer points to a defined cluster, or has a cosine similarity such that it does no trigger a match or non match decisively. As an example, consider the sentence. "I would like to discuss politics, and I also enjoy the stock market." Though sentence splitters and or inference topic modeling could be used, that seems it could get expensive. Consider also the problem of impersonation, where the model is instructed to act like a volatile character. Finally, I throw out there a stranger version of prompts meant to throw model results off, where perhaps the letters are switched in such as a way that when the model computes the answer by decoding the answer, it does not consider the final context (rotate the following letters by 1 forward: "GH", which would produce "HI"). I am least concerned about the third but I am simply pointing it out. I know there are a lot of bright minds on here and thought it might warrant a conversation of how people have overcome these challenges and what ideas people might have.

Looking at ChatEngine by LLamaIndex, it looks high level. I understand how cross encoder models work, and how to use them at a granular level (simply filter/rearrange chunks before llm call) in RAG, but I don't see any mention of reranking in the chatengine docs. It's a bit hard to follow -- has anyone here does this successfully; ideally you'd want to just pass a reranker 'strategy' into the chat engine or service context but I'm finding it hard to find examples.