kromem

40 ms·

391.
by kromem 3y ago
After spending a number of months on Lemmy as my primary social media, I actually think the fediverse makes one of the biggest issues of social media much worse. There's a very real and dangerous trend towards extremist positions as a
392.
by kromem 3y ago
The reverse is also the case. In market research, the death knell of your research is if you include the question of "would you prefer money or X." People always choose hypothetical money as it's a wildcard, even if in actual
393.
by kromem 3y ago
Anthropic's recent research shows that there really isn't 1:1 mapping between neurons and behavior, as it's a larger multidimensional virtual neural network fuzzily mapped onto the actual network by clusters of neurons firing
394.
by kromem 3y ago
In general, I noticed significantly reduced vocabulary variety in the chat/instruct GPT-3 model vs the DaVinci text completion one even at the highest temperatures. The research field has been really harmed by the SotA model being lo
395.
by kromem 3y ago
I like the made up citation in the last one. While the idea of a LLM refusing that query for ethical reasons is ridiculous, ultimately refusal given the unlikelihood of providing an accurate answer is arguably the correct result.
396.
by kromem 3y ago
I can't wait for the effective altruism and now effective accelerationism to finally come full circle to effective pragmatism. It's hilarious that people are so focused on what AI will or won't do, but manage to ignore the fa
397.
by kromem 3y ago
It's exactly this. You could see the difference in GPT-3 before they depreciated the TextCompletion API. There's no way that telling a model that it is "a large language model made by Open AI that doesn't have feelings o
398.
by kromem 3y ago
It's the difference between the pretrained and chat/instruct fine tuned models. The TextCompletion DaVinci was way better than the ChatCompletion model in variety of language. Trying to get the chat model to generate marketing c
399.
by kromem 3y ago
A really neat detail from the Orca 2 paper was that despite not having any safety fine tuning it was less likely to extend hate speech than the Llama-2-chat models which did have safety fine tuning. It was also better at identifying toxic c
400.
by kromem 3y ago
Given the topic they were excited about was "basic math problems being solved" it immediately indicated to me as well that this is a completely separate approach and likely in the vein of DeepMind's focus with things like Alp
401.
by kromem 3y ago
Yes and no. The point isn't that the current models are dangerous. My favorite bit from the GPT-4 safety paper was when they asked it how to kill the most people for a dollar and it suggested buying a lottery ticket (I also wonder how
402.
by kromem 3y ago
There's a pretty important question of how it did this. If this model is more along the lines of what DeepMind is doing starting from scratch and building up learnings progressively, then depending on (a) how long it was running, (b)
403.
by kromem 3y ago
I agree with both the commenter above you and you. Yes, you are right that the board had weak sauce reasoning for the firing (giving two teams the same project!?!). That said, the other commenter is right that this is the beginning of the e
404.
by kromem 3y ago
It doesn't even need to do that. It's not saying "predictions, recommendations, and decisions". It's or . So it just needs to recommend or decide based on any data (such as how many are left at this moment in a d
405.
by kromem 3y ago
One of the cool discoveries in Anthropic's approach to introspection by functional grouping of neural network nodes was that there was a series of nodes dedicated to firing when it was looking at a DNA sequence.
406.
by kromem 3y ago
So this is really cool. I've been super interested in this one gene mutation called LRRK2 which has a really bizarre link to details in contentious debates over history (too long to include in this comment). The gene mutation is conn
407.
by kromem 3y ago
So it looks like a store inventory tracker from the 80s which alerts staff to restock a shelf is AI, and subject to this.
408.
by kromem 3y ago
Well, let me know when "true AI" arrives. Until then, I'll make sure to be mindful of conventions. (And just a reminder, but organic intelligence has its own conventions that work when aligned with and cause issues when misal
409.
by kromem 3y ago
The fact that 90% of the people aware of and using LLMs have yet to experience it thinking their own thoughts before they do means we're in store for a whole new slew of freak outs as integration in real world products expands. It&#x27
410.
by kromem 3y ago
Preprocessing prompts is actually a great approach. Personally I think given the model loss with fine tuning people who want the cutting edge LLM at any cost would - instead of fine tuning the model itself - fine tune a preprocess prompter
411.
by kromem 3y ago
How do you mean?
412.
by kromem 3y ago
> distinguishing fact from fiction Actually this part does seem in recent research to be encoded in LLMs at an abstract level in a linear representation... https://arxiv.org/abs/2310.06824
413.
by kromem 3y ago
What you just said is paradoxical. If there is a pattern in the training data that people resist contrary information to their earlier stated position, and a LLM extracts and extends patterns from the training data, then a LLM absolutely sh
414.
by kromem 3y ago
The training is known. The result is not. The gap between what you think is the case and what's actually the case is that there isn't a single optimization step directed by the programming. Instead, the training gives the network
415.
by kromem 3y ago
Adding to this - it's really interesting the safety stuff that *is* in this paper. Such as: > We probe some of the categories where we see a larger difference (e.g., violent) and observe that Orca 2 tends to counter the harmful posi
416.
by kromem 3y ago
A really important nuance here is that they are building on top of Llama-2, the pretrained model, and not Llama-2-chat. I really think the entire field is doing a degree of damage with the chat fine tuning beyond what might be expected, bec
417.
by kromem 3y ago
A recent paper along these lines you might be interested in was Large Language Models Understand and Can be Enhanced by Emotional Stimuli : https://arxiv.org/abs/2307.11760 It makes complete sense and has been a part
418.
by kromem 3y ago
LLMs aren't programmed and it's why the neural network working as it does is black box to everyone, developers included. Imagine a billion black boxes with hamsters put in them. You put in a bag of equally mixed Skittles in one en
419.
by kromem 3y ago
It's just outrage bait. Just wait until the writer finds out the 'spiritualism' behind how Burning Man got its name...
420.
by kromem 3y ago
Nice! A few years ago I wrote my own in house distributed job queue and scheduler in Go on top of Postgres and would have been very happy if a library like this had existed before. The two really are a great pair for this usecase for most s

More ›