HN user
https://jimmyr.com https://www.linkedin.com/in/jimmyrcom/
Looks promising. I found a music generator called heymusic.ai and was considering subscribing, the songs were fun to make with the kids, but then it disappeared off the face of the earth with no updates anywhere online. I'm cautious of subscribing to AI services, lots of GEN AI startups popping up and it's not easy to make a profit
It doesn't sound like the approaches are incompatible. You can use minhash LSH to search a large set and get a top-k list for any individual, then use a weighted average with penalty rules to decide which of those qualifies as a dupe or not. Weighted minhash can also be used to efficiently add repeats to give some additional weighting.
Pretty amusing the old AI revolution was pure logic/reasoning/inference based. People knew to be a believable AI the system needed some level of believable reasoning and logic capabilities, but nobody wanted to decompose a business problem into disjunctive logic statements, and any additional logic can have implications across the whole universe of other logic making it hard to predict and maintain.
LLMs brought this new revolution where it's not immediately obvious you're chatting with a machine, but, just like most humans, they still severely lack the ability to decompose unstructured data into logic statements and prove anything out. It would be amazing if they could write some datalog or prolog to approximate more complex neural-network-based understanding of some problem, as logic based systems are more explainable
Isn't this how most vector search backends work, cosine similarity sorting against whatever the vector embeddings returned? How is this different?
Thanks!
Looking forward to an 8bit instruct version on llama.cpp to try out problems with the insane context length.
It would be interesting if all these models were finetuned on basic datalog which is a very simple language. That way they could demonstrate their logic/reasoning capabilities as well as ability to learn from mistakes and iterate.
Sounds like an implementation for how to managing scopes using stack frames with some state machine like elements to it
If they add fstrings, some type of easier map get/put syntax eg #map.value = 1, and maybe a shorter hand fun syntax, then Erlang feels like it's gotten all the conveniences of python. Amazing how far things have come
Sounds like it's just a Haskell thunk but as a probability wave
Lazily evaluated until there's a probability it has to interact with something. Since you can never really see the value of the actual function, but only see what it looks like when it's forced to evaluate a computation in some context, an interaction, you can never get a precise definition of the function
Go to settings -> extensions and make sure youtube is enabled Maybe it only works for gemini advanced
Some guy will put videos of a patient getting lasik/*, records all the movements of the surgeon as input, shove it into a transformer model and there will be machines which do it 10x better than humans in 5 years. Might as well wait
Racket made a prolog implementation that is currently used in the compiler, or at least I thought I heard that in one of their talks. Might be worth trying, since its model is a language for making languages
They're sending people a bunch of invoices without any due dilligence. You're going to get companies that double pay because they're not sure/don't have time to review licensing contracts. The business model is very scam inspired.
Scammers are sending me geeksquad invoice emails every week. Scammers sent fake invoices to GOOG and META and receive 100m for it [1]. This feels no different, just scan for sites using your font and send invoices frequently whether they have a license or not. It's a very hacker/scammer type of approach to making money, or I guess you could call it a private equity font-industry-vertical 10x secret sauce.
https://www.npr.org/2019/03/25/706715377/man-pleads-guilty-t...
Twitter and social are also listing tons of these "You've heard of chatGPT but these are the next ...", half these posts look like clones of each other but they all seem to have a huge amount of shares.
I remember del.icio.us from a long time ago was chocked full of top 100 lists, for example giant directories of free OCW courses. Stuff no one ever goes through, but everyone feels like they have to bookmark to go through it later.
95% of the professional asset managers, for example those in wealthion, blockworks, paid macro people like 42macro and themacrocompass, nearly all of them, said stocks would most likely touch 4300 on the S&P and probably cycle down again to prior lows, as manufacturing PMIs, housing, etc weaken, as the long and variable lags from Fed tightening, credit environment, start to hit.
Instead, right after the debt ceiling, there was a massive short squeeze, parabolic AI tech pump, and we're at 4567 on the S&P.
As it turns out the people just trading off momentum, technical analysis, and liquidity expectations, did way better than those betting on certain industries to go down. Sure, it can still go down, but there are plenty of money managers, macro experts, that looked at the big picture based and made data driven decisions based on historical data, and still got completely burned because they were trading against the technicals (massive upward momentum since after October).
The pre-prompting length will grow more and more as more liabilities in the responses are uncovered. I would imagine the more the pre-prompting grows the more attention is diverted to the rules rather than the user prompt, and the less reasoning available.
I wonder if they'll start using LLMs while ingesting new data. eg asking the LLM if the content is helpful, cites sources, respectful, positive, not-thin content, common or often duplicated content, etc etc, before each content import.
Well open AI raised eye brows by crawling the internet and using everyone's data to make a commercial product
One day some new startup will train on all of libgen and torrent networks, but it will be very hard to prove. You'll keep getting these gaps up in questionable morality and legality, and even openai will complain about playing fair
This is great, love it!
Crawling sites to index the FAQ's and knowledge bases, into the vector search, isn't as intimidating as it sounds, at least on linux systems. Sometimes a thin wrapper function over plain old wget will get you 99% of the way
wget -rnH -t 1 --waitretry=0 'https://{{domain}}' -P '{{domain}}'Protect yourself from sophisticated bots, which can mess up your traffic reports
Deep forgetting
Lots of people using LLMs to make chat bots from their existing datasets: customer service troubleshooting, FAQs, billing, scheduling. Being able to upload their own pdfs, spreadsheets, docx, crawl their home page, lets the chat bot become personalized to their use case. While you could locally query your own vectordb before prompting, people buy paid service so they won't have to manage any of the technical details.
If people can drag and drop some files from their nas, you parse them with apache tika or similar https://tika.apache.org/ , they can start using personalized branded bots. It also lets you do things like refusing to answer, if the vector database returns nothing and the use case requires a specific answer from the docs only (not the llm to make stuff up).
It's difficult to compete. A small business might answer 10,000 requests to their chat bot. The options are
- Pay openai less than $50mo
- Manage cloud gpus, hire ml engineers > $1000/mo
- Buy a local 4090 and put it under someone's desk, $no reliability +$1500 fixed
Any larger business will need scalability and you still can't compete with openai pricing.
Maybe one of you startup inclined people can make an openllama startup that charges by request and allows for finetuning, vector storage
Anyone know how milvus, quickwit, pinecone compares?
I've been thinking about seeing if there's consulting opportunities for local businesses for LLMs, finetuning/vector search, chat bots. Also making tools to make it easier to drag and drop files and get personalized inference. Recently I saw this one pop into my linkedin feed, https://gpt-trainer.com/ . There's been a few others for documents I've found
Nope nope, wouldn't want to compete with that on pricing. Local open source LLMs on a 3090 would also be a cool service, but wouldn't have any scalability.
Are there any other finetuning or vector search context startups you've seen?
The stock trades like some crypto alt coin from the promised land
I wonder if the pre-prompting part was increased as part of the trust/safety effort. For example they increased more examples of the types of things not to say, attached to the user prompts. That would decrease the amount of reasoning it could give the actual prompt, as it has to logically make sure each statement complies with the prior rules, and it also decreases the context length.
Looking at producthunt there's tons of new SEO AI tools, things like surfer for example. They try to auto-generate the target keywords, blog sections etc while still trying to adhere to quality checklists https://developers.google.com/search/docs/specialty/ecommerc... . Funny how posts that speak from experience and show authority may have higher quality, but chatGPT will flat out protest if it's asked to give it's personal experience, giving away who is using the bots.
Looks like it's going to be a long battle between the bots and the search engines.