Palmik

24 ms·

421.
by Palmik 3y ago
Advanced code generation, the kind you would need for something like AutoGPT or other agent-like system, is definitely reasoning and GPT-4 or Claude 2 will dominate. But most of the code generation from GitHub Copilot or other systems (like
422.
by Palmik 3y ago
It's great that the costs are coming down, but it's still not economical for many activities. The fact that some things had extreme margins before, and now they have less extreme margins, isn't a really good indicator.
423.
by Palmik 3y ago
I am totally with you when it comes to reasoning, or when it comes to generally / universally capable models. But many problems don't require reasoning, and are very concrete / specific.
424.
by Palmik 3y ago
Fine tuned gpt-3.5 is ~8x the price compared to regular gpt-3.5. It might be a good way to collect better training data though. :)
425.
by Palmik 3y ago
Sure, there are many ways one can be accidentally inefficient, some examples: Batching: Your throughput (and thus cost per token) will be much worse if you don't do batching (meaning running inference for several inputs at once). To do
426.
by Palmik 3y ago
Where fine tuning shines imo: Things that require consistency: e.g. you want the chat / output to have certain "personality", consistent level of conciseness or formatting. Things where examples are hard to fit into prompt: e
427.
by Palmik 3y ago
Good point -- hopefully the quality impact is still worth it, remains to be seen. Agree on the size -- hopefully something they will keep in mind for future models.
428.
by Palmik 3y ago
Why this is interesting: To my knowledge, this is the first time you are able to run the largest Llama, at competitive speed, on a consumer GPU (or something like A40). On many tasks, fine-tuned Llama can outperform GPT-3.5-turbo or even GP
429.
Exllamav2: Inference library for running LLMs locally on consumer-class GPUs(github.com)
322 points by Palmik 3y ago|125 comments
430.
by Palmik 3y ago
None of these concerns seem very specific to AI startups (*). Like the defensibility and competitive advantage: These concerns exist with most (all?) non-AI startups as well, and they are equally valid. See countless examples of incumbents
431.
My time with Rails is up (2016)(solnic.codes)
39 points by Palmik 3y ago|47 comments
432.
Roblox's game building AI assistant(theverge.com)
1 points by Palmik 3y ago|0 comments
433.
by Palmik 3y ago
In general, I think it's best to avoid stereotypes like this, unless they aren't just stereotypes but a statistic. And even then, interpreting a statistic can often be tricky.
434.
by Palmik 3y ago
This comment might be interpreted as offensive on so many levels. :) One of them being the implication that, on average, nationals are bigger slackers than foreigners.
435.
by Palmik 3y ago
Significant amount of tech workers were affected, at least at Google.
436.
by Palmik 3y ago
Privacy policy summaries + q&a would be a nice public service. Here's summary of this one: > Personal data collected includes user names, text or image prompts, IP addresses, contact information, and more. The data is used for v
437.
by Palmik 3y ago
If you are interested in on demand summaries, you can use https://dstill.ai/hackernews to summarize any story -- it summarizes not only the link, but also the post & comments. Previously posted about this here: https:&
438.
TIME100 AI(time.com)
9 points by Palmik 3y ago|0 comments
439.
by Palmik 3y ago
At Google, there are no branches, so typically development indeed happens under some experiment flag. Binary, configs and experiment flags are on separate deployment schedules, and often you need to coordinate your release across different
440.
by Palmik 3y ago
If you find this interesting, checkout also "Mass Editing Memory in a Transformer" [1] and "Locating and Editing Factual Associations in GPT" [2]. [1] https://memit.baulab.info/ [2] https://ro
441.
by Palmik 3y ago
You likely won't see performance (speed) benefits if you have just tens of thousands of vectors and of course brute force (vector matrix multiplication) will have better precision & recall. Vector databases make sense when you have
442.
Google Cloud and Nvidia Join Forces to Power AI Computing(wccftech.com)
1 points by Palmik 3y ago|0 comments
443.
by Palmik 3y ago
According to Clement (the HF CEO) they are close to covering their costs, and have 10years runway: https://twitter.com/ClementDelangue/status/16947653579682861...
444.
by Palmik 3y ago
The best model, Unnatural Code Llama, is not released. Likely because it's trained on GPT4 based data, and might violate OpenAI TOS, because as per the "Unnatural" paper [1], the "unnatural" data is generated with t
445.
by Palmik 3y ago
There's also https://github.com/windmill-labs/windmill and https://github.com/activepieces/activepieces depending on which parts of retool and tooljet you care about.
446.
Announcing Ideogram AI(ideogram.ai)
3 points by Palmik 3y ago|1 comments
447.
by Palmik 3y ago
Nice work! Curious how you see the differentiation from similar products, like https://www.trudo.ai/ , https://promptitude.io/ , https://www.fixie.ai/ , https://www.baseplate.ai&#x2
448.
by Palmik 3y ago
Can you clarify about resveratrol? My understanding is that the original study had issues and wasn't reproduced. This video has some details: https://youtu.be/JAFnD27ffqE I'm holding more hope for rapamycin, which
449.
by Palmik 3y ago
This looks quite interesting, congrats on the launch! Reminiscent of fly.io. Is it a direct competitor, or is there a major twist to it? How do you handle apps composed of multiple services, if there isn't a configuration? The pricing
450.
What Types of Questions Require Conversation to Answer? AskReddit Study(arxiv.org)
1 points by Palmik 3y ago|0 comments

More ›