69 karma · joined September 7, 2013
Peter did done great ground work here and as things transitioned to neural nets, also peter appears randomly in my Chromecast with some nice photos he took lol
Lots of good docs have you visually doing the setup/work, then you realise you can download/see a bunch of code for all the actions you did (e.g. everything UI doable is programmable and the tools taught you all the concepts), it minimises the actual programming, it is really hard/time consuming to mentaly translate API docs to code.
If the open source community then other company's didn't follow along with them then we would still be being price gauged in a monopoly situation, hopefully same thing happens to dalle soon
I'd be interested in how to convert models without retaining, or minimal retaining, I'd have thought it would just work with iterative mask tokens at least for models trained with MLM
### migrations/user_add_free_credit.sql adds free credits field to user table
This space needs some competition
With text-generator.io your only charged per request + a free tier so it's much more predictable and and you can generate lots of variants in one request. It's also API compatible so only a one line change, it doesn't do the dalle stuff though I don't think the competitors or open source image generation has caught up yet but won't be long
I'm confident in selling it for 25k per year to businesses because:
* there's a one line migration in from OpenAI
* Self housting will save the companies entire OpenAI bill
* OpenAI doesn't have some features I do like autocomplete or generating a set number of sentences
* Generating multi lingual text and code is a hard problem to solve yourself, specifically it's a larger than 25k development problem, so will cost more if you want to try solve it yourself right now, so businesses would rather pay some money to have it now which also reduces risk for them.
With a lot of experience in a field and hard work it's very possible to create something others can't feasibly do, or do for the price you do
Similar to some answers about including a few example questions/answers prefixed with Q: and A: to put the model into a more factual question answering mode.
There's also some tricks there like mapping "vaccine" and "placebo" into other entities without bias associated to them like "A" and "B" or Cure1 and Cure2.
Try the models at https://text-generator.io not sure if that will have the same biases or not but if not then you'll at least see massive cost savings vs OpenAI
They mostly helped me learn/grow on to better roles and learn more about creating businesses and machine learning.
My main income so far was mostly powered by house flipping/stock (canva)/work but i recently left to persue running my own AI business.
I now made/run https://text-generator.io an API for devs to generate text/code which recently became profitable.
I find with careers in general it's best to be the fearful one feeling like the perpetual imposter around greats, taking too many risks, moving around what you do a lot to build your skills, i received lots of negative feedback about ideas before e.g. "its impossible to predict the stock market", "if your predictions are so good then why arent you just keeping it to yourself", "why compete with Github copilot on generating code?" "arent they so good already", "isn't it too ambitious", "why compete with OpenAI", turns out there's lots of potential customers who are in much different situations than ones giving negative feedback, receiving the negative feedback helps you feel scared and know your doing something against the status quo.
I made some massive "blunders at face value" like leaving Canva and blowing lots of potential stock to go learn about using Machine Learning for generating text at a startup, blowing lots of money developing stock trading algorithms etc...
While taking the appropriate amount of risks to have a fulfilling life, you'll find it feels like a rollercoaster filled with ups and downs that may feel like epic fails but are really epic lessons.
I have no idea why app.nz is blocked, its a shame because there's some free products i've been building there to help kids read (readingtime.app.nz) and their creativity (joydrop.app.nz).
Maybe the app.nz domain was used for bad things before i bought it, that's my only stab in the dark at why they might have blocked it. Moving domains could help in these situations where you have been false positive flagged then obliterated by some internet police style company e.g. FAANG or stripe etc.
Note when generating text try https://text-generator.io which will save you some when generating text or code Vs open ai
* checkout mosaic ml, it has some recent improvements like Alibi that are important https://docs.mosaicml.com/en/v0.3.1/method_cards/alibi.html
* Huggingface trainer
* Pytorch lightning (still has some great training improvements like LR find and saving top k best checkpoints ect)
* Deepspeed/megatron/bloom
For inference:
try a few alternatives to OpenAI like https://text-generator.io/
or hosting your own network/huggingface pipelines (can use a feature extractor pipeline and a text generation pipeline to support getting embeddings and generating from the same one model on the GPU)
They just seem to loop around a lot i don't know why... i think they trained on datasets with duplicate content or lots of repeating characters.
Check out https://text-generator.io which is orders of magnitude cheaper than GPT-3 and still makes creative writing/code autocomplete without too much looping issues that youd see with OPT models.
For that repeating you can dial up the repetition penalty or N to generate more sequences in a single request (and you are only charged by the request not by characters/tokens which helps), often generating N results is much more creative than generating a long result as the generated output gets fed into the input when generating long text and often causes that kind of repetitive looping.
Some tactics to mitigate that repetitiveness are:
* Dont use OPT...
* repetition_penalty/retries/seed
* Generate N results and combine instead of doing one big generate as you dont know how many results you'll really get and they are less creative/likely contain shared info/repetitiveness.
* Creative input prompts
There's probably other creative ways of getting variety by splitting the generation into different calls with higher repetition penalty as you get longer or looping detection etc. Its easy to detect repetition in the structured case like chat but hard when doing longer text generation/creative writing
Also checkout https://text-generator.io ... not SOTA.. as its not using hyper large language models, but does a excellent job for basically price/performance vs OpenAI which is very pricey. Also generates code without having to go on a waitlist for codex.
https://text-generator.io/playground?text=Write+a+story+abou...
"""
Write a story about Ash and Pikachu commanding an army of Pokemon. They're attacking Team Rocket and enemy Pokemon from the land and ocean.
A:
I think you can do this with PokeBalls.
To make it more interesting, there are many ways to get them like making your own one by picking up some random item in Pokémon City or even getting free ones on Amazon (or any other website selling stuff). You could also go for something unique such as taking part in Battle Royale where all players have exactly 2 pokedexes each containing their favourite pokémon. Another thing that makes these games so much fun is how they keep evolving over time - sometimes adding new features while at times removing old functions which will change what happens next when you're playing online against others using different versions.
"""
Heres attempt number 2 which i was inspired by the epic "it was a dark and stormy night" intro that gpt-3 made up, so i just added "Once upon a time" its good to prompt these things with some info to make sure it knows its not doing question answering like with the above prompt
""" Write a story about Ash and Pikachu commanding an army of Pokemon. They're attacking Team Rocket and enemy Pokemon from the land and ocean.
Once upon a time, there was no Pokémon world.
There were only three creatures in this universe--
One who loved to fly,
A girl with wings...
And one person whose name I can't pronounce.
(Note: This is not quite as long or detailed here.)
The third creature had great powers that he could use for good
(like flight) but also bad (such as flying into flames).
"""
Quite hilarious there...
its SOTA!
Also checkout https://text-generator.io ... not SOTA.. as its not using hyper large language models, but does a excellent job for basically price/performance vs OpenAI which is very pricey. Also generates code without having to go on a waitlist for codex.
That is to say PII you send to OpenAI as their customer might be leaked whereas https://text-generator.io persists nothing of the customer input data and doesn't train on it.
In terms of whats in there right now though its likely trained on the pile though so you'd probably already be in there dataset too before being a customer, basically Reddit is in there/hacker news/Github i believe too, same thing if your authoring pubmeds ect so its hard to keep anonymous forever
the same querying works but it will save you a load on gpt-3 cost https://text-generator.io/playground?text=SQLite+create+a+pa...
codex is technically free if you have access but soon it wont be so keep text-generator.io in mind there too as it can generate code from descriptions etc too at a wildly more competitive price.
Had to use a bit of the "Create Table" syntax, but it did a create table if not exists which is nice... it missed the content field though
Would love to hear how this works.
Check out https://text-generator.io which is much cheaper than GPT-3 and still makes creative writing/code autocomplete.
You can dial up the repetition penalty or N to generate more sequences in a single request (and you are only charged by the request not by characters/tokens which helps), often generating N results is much more creative than generating a long result as the generated output gets fed into the input when generating long text and often causes that kind of repetitive looping.
So some tactics to mitigate that repetitiveness are: * repetition_penalty/retries/seed * Generate N results and combine instead of doing one big generate and trying to get 40 results and ending up with 10 similar ones * Creative input prompts
There's probably other creative ways of getting variety by splitting the generation into different calls with higher repetition penalty as you get longer or looping detection etc. Its easy to detect repetition in the structured case like chat but hard when doing longer text generation/creative writing
https://text-generator.io/blog/over-10x-openai-cost-savings-...
Open AI is massively expensive/this has a free tier of 100 requests forever and is priced at .01c per request instead of based on input+output tokens, it should be over 10x cheaper, i'm willing to lower pricing to make sure its 10x cheaper if anyone doesn't see 10x cost savings after switching over :) Let me know how it goes and enjoy some cost savings!
Also the API saves nothing of what you send to it so it's very secure in that regard.
Let me know what use cases you have for it, exited to see the technology get applied to benefit us all a lot more as pricing becomes reasonable.
I developed some games like https://github.com/lee101/wordsmashing in a fairly simplistic jquery style i developed where i purposefully avoid using features of JS like "this", it trips most developers up when you pass functions around and the reference to "this" gets lost.
Agree with lots of the comments here about complexity of both JS (new es module compatibility issues etc) and JS frameworks that are limiting what people can actually build these days from cognitive overload and bugs.
a tonne of technologies (like polymer.io/web components) died in a sea of complexity and much technology we use is going to go the same way unfortunately.
Another gripe i think unpopular but true is that we as coders prefer complex things like static site generators like Jekyll/ghost etc, normal people can use a CMS and get stuff done without having to resolve packaging conflicts and SSL issues, we have to resist the urge to pretend coding/setting up complex software is easy because its not... also Kubernetes...
They gave us a stock rails CMS/static site generator that did the frontend part.
almost full rewrite as we where a Python shop and there was a mess of JQuery pubsub involved so you wouldn't know what was happening in what order or if there would be a bug because something wasn't subscribed to this event bus in time. almost goto level.
We met with the external company a few times and they made it clear to us they wheren't getting paid to fix tech debt or make the handover process smooth for us just that they where getting paid to make it look like the design.
This is actually a better state than the other projects i can think of, more because its a small project and the levels of technical debt are thus lower because theres not as much code/complexity, not that the code that is there was good.
All the tests where failing (very flaky by design, using browser driving to check for very specific parts of text on the page that had long been replaced), we also deleted all their tests/writ our own aswell as the almost full rewrite.
This just really messes with the concept of making peoples money less valuable than other peoples by blacklisting addresses, wont work id say for cryptoNote currencies ...
plug: checkout https://bitbank.nz for ai market forecasts