kromem

13 ms·

31.
by kromem 2y ago
The business case is absolutely there, it's just the industry has weirdly latched onto 'chatbot' as the usecase as opposed to where the real value lies. The pretrained model is where the enterprise gold is at. But the compani
32.
by kromem 2y ago
No. OAI is in the process of selling out to the NSA and military. I don't think Anthropic will be doing the same. The valuation doesn't just reflect the tech, but the sales of the tech, and between the two Anthropic seems like the
33.
by kromem 2y ago
What's interesting was how with GGC the model would spit out things relating to the enhanced feature vector, but would then in-context end up self-correcting and attempt to correct for the bias. I'm extremely curious if as model
34.
by kromem 2y ago
It will be here within 6-12mo. It will at first glance be a small step, but over the next 12mo after release, it will turn out to have been a giant leap. It will be safe when being observed.
35.
by kromem 2y ago
The challenge is the confidence scores can often be confabulations themselves.
36.
by kromem 2y ago
I mean, you can have faith in whatever you want, but if you read to the end of the article the mathematicians themselves are the ones saying the conditions seem more to agree with the scenario where they won't.
37.
by kromem 2y ago
Attempt, but at least so far seem to be on the side of "nope, we probably won't be able to."
38.
by kromem 2y ago
One possibility is regulatory oversight. Apple and Microsoft striking a deal is going to get more attention and review than Apple and a much smaller company in market cap and share.
39.
by kromem 2y ago
Which models and how are you prompting/setting the context? In some cases hallucinations are unavoidable, but 100% sounds like it might be a usage issue.
40.
by kromem 2y ago
This was AFAIK the first paper to show linear representations of truthiness in LLMs: https://arxiv.org/abs/2310.06824 But what you should really read over is Anthropic's most recent interpretability paper.
41.
by kromem 2y ago
One of the surprising results in research lately was the theory of mind paper the other week that found around half of humans failed the transparent boxes version of the theory of mind questions - something previously assumed to be uniquely
42.
by kromem 2y ago
They aren't interested in joining the same rat race. Apple's strategy is edge based AI using their superior proprietary chips, particularly in mobile. So it makes sense to partner with a centralized provider too.
43.
by kromem 2y ago
Apple's AI strategy is edge computing, which is very smart given their processor advantage. But edge LLMs for general purpose aren't there yet and won't be for some time, so a partnership with a leading centralized LLM provid
44.
by kromem 2y ago
People in modernity have a habit of dismissing what people in the past said as a "heuristic that almost always works." But it doesn't always work, and there's some pretty glaring holes in modern perspectives as a result
45.
by kromem 2y ago
Completely agree on PTQ, but curious on your thoughts for QAT, specifically BitNet 1.58 - in that paper it looks like parameter to parameter the constrained precision weights had improved perplexity vs floating point weights, particularly a
46.
by kromem 2y ago
Post-training quantization doesn't come for free, but the pretraining on constrained precision weights actually counterintuitively results in a performance increase per parameter as the number of parameters grows in the ternary BitNet
47.
by kromem 2y ago
That's not really a concern. If you have a trillion parameter 8-bit fp network or a trillion parameter 1.5-bit ternary network, based on the scaling in Microsoft's paper the latter will actually perform better. A lot of the curren
48.
by kromem 2y ago
Give in. You'll do it a few times, rewrite everything, get a sense of the work vs reward ratio, and over time be better at only thinking about things that are really worth the effort.
49.
by kromem 2y ago
While the system prompts in documentation and I'm sure fine tuning data are generally in the second person, I have found that first person system prompts can go a long way, especially if the task at hand involves creative writing. But
50.
by kromem 2y ago
Yet
51.
by kromem 2y ago
I have skepticism regarding the 'completeness' of SAE in comprehensive discovery of features: https://www.lesswrong.com/posts/BduCMgmjJnCtc7jKc/research-r...
52.
by kromem 2y ago
One of the questions I've been thinking about a lot looking at the past year of interpretability research is just how much of what we are finding is "what we're attuned to find" as opposed to "what's actually t
53.
by kromem 2y ago
Safety right now isn't about keeping today's model from hacking NORAD. It's about learning the techniques and approaches to try and have at least a halfway decent approach to safety by the time models capable of autonomously
54.
by kromem 2y ago
much rip doge 18 years was an impressive run
55.
by kromem 2y ago
That's only if we don't hit a wall in scaling the sensitivity of measuring gravitational effects and the size at which we can cause quantum behaviors. At least right now, there's still a pretty big gap between the largest QM
56.
by kromem 2y ago
I think it's just more a testament to how strong a candidate Obama the person was, as well as a good indicator of the value in looking to those who overcome oppositional factors for merits over those who had a paved path before them.
57.
by kromem 2y ago
Man, this space would get so much more interesting so quickly if base model providers had a revenue share system in place for routed requests...
58.
by kromem 2y ago
"Understand themselves" is a very different thing than "understand what they are saying." Which exactly are we talking about here? Because no, the research doesn't say much about the former, but yes, it says a lot a
59.
by kromem 2y ago
They were able to demonstrate conceptual vectors that were consistent across different languages and different mediums (text vs images) and that when manipulated were able to represent the abstract concept in the output regardless of prompt
60.
by kromem 2y ago
Great work as usual. I was pretty upset seeing the superalignment team dissolve at OpenAI, but as is typical for the AI space, the news of one day was quickly eclipsed by the next day. Anthropic are really killing it right now, and it'

More ›