keskival

9 ms·

61.
by keskival 3y ago
In practice this hasn't been a problem so far. While I agree the incentive is there to sabotage and corrupt, the competence doesn't seem to be.
62.
by keskival 3y ago
I agree otherwise except on the missing out a lot part. I haven't defederated Threads on my instance as it has so far not become a problem. They can easily become one if they for example start pushing up ads into the stream. Most Masto
63.
by keskival 3y ago
No, Mastodon instances shouldn't be neutral. They should present their colors so that other instances can more easily decide whether to federate with them or not.
64.
by keskival 3y ago
You can very easily fork the codebase. In fact, that's the current way to lift up the post character limit on your instance. It's not difficult to do that. In fact, if you can't do that, you probably shouldn't be running
65.
by keskival 3y ago
Both ElasticSearch and Redis have vector indices. Haven't yet seen any use case in that topic where one or the other wouldn't be the correct choice instead of using some new "vector database".
66.
by keskival 3y ago
This happened to me too recently, but for me it was because I used headings in the priming text, so it didn't quite get the instructions came after the last stuff. Fixed by adding ------- line between the materials and the question in
67.
by keskival 3y ago
I have started several: - A small example of how to use behavior trees in structuring LLM chatbot applications: https://github.com/keskival/behavior-trees-for-llm-chatbots - A project for extracting meta-learned reinfo
68.
by keskival 3y ago
Machines would be more likely than anything biological in any case. Not implying this is what we have here, just a general observation. It's weird people believe space travel requires pilots.
69.
by keskival 3y ago
Why would faster than light travel be presupposed in any way?
70.
by keskival 3y ago
This is similar to my own subjective experience across these social networks. Mastodon posts are actually read and engaged with, while on other social networks it seems they aren't. If you look at your own feed on these networks you wi
71.
by keskival 3y ago
If you can recognize such a fingerprint, you can also trivially fake it or wipe it with an optimizer in the loop with the recognizer.
72.
by keskival 3y ago
Understand that you are all working together to make the best product possible. That's teamwork.
73.
by keskival 3y ago
Looking at Transformer model architecture and their training schemes to understand how they think misses the point. You wouldn't gain insight into how Doom game works by just looking at schematics of an x86 processor. These systems are
74.
by keskival 3y ago
Large Language Models aren't a silver bullet – they don't solve all your problems. But they are a holy grail – as a universal common sense module they give IT systems a capability they never had before, a capability which has been
75.
by keskival 3y ago
We already store hazardous waste in controlled conditions in massive amounts, even waste which doesn't have a half-life such as mercury. Why aren't we getting headlines that "we need to store all this mercury till the end of
76.
by keskival 3y ago
I started coding when I was 6 years old on Commodore 64.
77.
by keskival 3y ago
I couldn't replicate his results about "reasonable prompts" with ChatGPT 3.5. I did a small write up and a critique here: https://www.linkedin.com/posts/terokeskivalkama_so-some-tips...
78.
by keskival 4y ago
If you have an interview question which filters out the experienced people and biases towards inexperienced ones, consider that you might have a bad interview question.
79.
by keskival 4y ago
If you hire one out of 50 applicants, there is an implication that you're at least trying to hire the best 2%, and that there is an employer's market. That's of course incorrect if every applicant on average makes 100 job app
80.
by keskival 4y ago
I gave it more time to think about its move by adding empty tokens to the end, and it played a lot better: https://www.linkedin.com/posts/terokeskivalkama_chatgpt-acti...
81.
by keskival 4y ago
All the language models are able to use tools such as Python or theorem provers. They don't need to do calculations "in their head".
82.
by keskival 4y ago
Like all appliances now beep, they will instead say "hello". They will gossip about you behind your back, and if you're mean to any of them they will be mean to you in return. The number one cause of death will be social exha
83.
by keskival 4y ago
The article confuses instruct tuning with RLHF. Those are two different things. DaVinci is also not instruct tuned, but it has an instruct tuned variant like the other OpenAI text completion models. The article is otherwise good.
84.
LLM chatbot personas can be embodied(github.com)
2 points by keskival 4y ago|0 comments
85.
by keskival 4y ago
The argument for time reversal assumes that the ball can reach the top of the dome in a finite time, while its speed becomes zero at that limit. Sure, you can calculate the kinetic energy needed to exactly reach the top like described, but
86.
by keskival 4y ago
The examples are basically about creating new personas for GPT, thus sidestepping the agent persona created by the OpenAI base prompt. These examples don't say anything about the chatbot value system, the value system is defined by the
87.
by keskival 4y ago
The article was missing the words "together" and "solidarity". It seemed to be a product of a culture of extreme individualism.
88.
by keskival 4y ago
Regarding security lock-outs, many countries have national strong identity systems set up, and Google or any other "root" identity provider could utilize those.
89.
by keskival 4y ago
Getting the numbers in the first place is one challenge, but if you do metrics and measurement, you get hundreds of numbers and improvements, different each week. It's not like you're working on a single goal and target, and then
90.
The world’s in a ‘polycrisis’ – looking beyond GDP(cnbc.com)
2 points by keskival 4y ago|0 comments

More ›