lumost

39 ms·

811.
by lumost 3y ago
internet was both a monumentally successful technology, and a bubble. The dot-com crash of 2000 saw most stocks give up all gains from the prior ~5 years, the surviving stocks took over a decade to recover their previous high.
812.
by lumost 3y ago
like having a very visible/active chairman of the board.
813.
by lumost 3y ago
Is there a good method to split out memory and CPU bottlenecks on a CPU? most folks simply optimize until they hit 100% and assume all CPU usage is created equal. Seems like there needs to be a tool similar to htop with utiization stats for
814.
by lumost 3y ago
It did not, health plan was covered by a separate charge through the university or through student's own continued coverage from parents.
815.
by lumost 3y ago
Anecdotally, US labor union relations have been strained for some time compared to international settings. Plausibly, this is in part due to the size of American labor unions compared to their international equivalents. The Resident Advisor
816.
by lumost 3y ago
Outside of hardware/implementation optimizations, and position embedding choice - has the SOTA transformer architecture evolved that much? Llama-2 code appears to be about the same as gpt-2.
817.
by lumost 3y ago
In the cloud, these chips will compete head to head with GPUs. If they are able to pull off a 10x price/performance win without excessive porting work… it’ll take off in a heartbeat.
818.
by lumost 3y ago
Startups are typically capex constrained, at least until series C. Clouds are favorable to capex constraints. Having personally tried to be cheaper than the cloud in 2013 through large hardware buys, negotiated contracts etc. I found that t
819.
by lumost 3y ago
It's incredibly difficult for a mature software business to justify infrastructure and tooling investments. This is why we think that startups are a haven for modern tooling and the largest legacy firms are ... well ... difficult. The
820.
by lumost 3y ago
The LLM is also trained to predict next token given a sequence of "real" tokens. This means that any case where the LLM predicts tokens that don't look like the "real" token string is Out of Domain. Children learn i
821.
by lumost 3y ago
Folks also underestimate how complex these libraries are. There are dozens of projects to make BLAS alternatives which give up after ~3-6 months when they realize that this project will take years to be successful.
822.
by lumost 3y ago
>But by-and-large, it targets existing use-cases, for the simple reason that you need training data to build the AI in the first place. This doesn't strike me as a true statement. People are embracing the term "AI" to repr
823.
by lumost 3y ago
I agree that oversight is needed - but I'm not certain people are the right approach to this oversight. Everything described above represents an undesirable software output for the given business. Why can't we make the LLM smart e
824.
by lumost 3y ago
I really dislike the current attitudes around ethics and AI. Tools should be 100% open about what they can and can't do. Building unethical software like deepfakes for porn is also unethical. Quoting of vague "ethics" risks a
825.
by lumost 3y ago
GPUs will be cheaper later. GPUs have still been tracking Moore's law, enterprise GPUs went from a niche product to being the hottest enterprise hardware overnight. Margins on H200's are massive from what I understand. I would bet
826.
by lumost 3y ago
I’d be curious if $20/mo is or is not enough to cover their costs. It’s opaque what their inference setup is, or what the final economics of LLMs will be. With 30-100 Billion param models becoming available which are in spitting distan
827.
by lumost 3y ago
Anybody remember the google duo demo?
828.
by lumost 3y ago
If they tried to focus on ads, then they wouldn’t have the talent to support the business. They probably don’t need 17 chat apps - but they can’t start saying no without having other problems.
829.
by lumost 3y ago
It probably depends on your objective. 352 lb is about the cutoff before you become a competitive weight lifter. Most people probably don't want to do that. Same thing happens in gaming, 95th percentile is probably about where you get
830.
by lumost 3y ago
This is almost certainly going to be used to generate actual pictures of real people in the nude etc.
831.
by lumost 3y ago
The gradient is the direction and magnitude of change for each model weight. The optimizer determines how to adjust the weight based on the gradient.
832.
by lumost 3y ago
Pretty great! Quick q for you when you update the docs. How does front end rendering work? Much of rust is deep into wasm - are you following that approach? or aiming for pure ssr of templates?
833.
by lumost 3y ago
Anecdotally, The field of ML suffered from a problem where there were more entrants to the field than available positions/viable work. In many industrial positions, it was possible to hide a lack of progress behind ambiguity and/o
834.
by lumost 3y ago
There are a number of dialogs you can have with the models where the model plays a persona. Eg “you are a librarian, please only give accurate answers or references or you will be fired” GPT3.5 will suddenly become extremely risk averse in
835.
by lumost 3y ago
I strongly doubt that the LLM club will remain exclusive for very long, or that the innovation in this space was restricted to 10k people. At a minimum, I’ve already begun using gpt-4 as a consultant for my fathers medical issues. I can rea
836.
by lumost 3y ago
My hunch, The history of open source has been that companies who have customers with massive customization requirements land on the open source side of the equation. Companies who don't view a component as core to their product often l
837.
by lumost 3y ago
As a fan of big science projects, I find this line of thinking unconvincing. The manhattan project cost roughly 56 billion in todays dollars vs. 149 billion for the Apollo program. The manhattan project effectively ended global great power
838.
by lumost 3y ago
7 figures is pretty great for an individual, assuming that there is at least a 50% margin. For a company it's a good start in SF. Scaling a business with no moat will quickly bump into margin compression and a race to the bottom. Not a
839.
by lumost 3y ago
I always found it strange that VCs were so focused on GPT for X companies. In my view, there is no reason why GPT for X isn't just GPT. Once the model is fine tuned for this task, there isn't anything terribly special about a give
840.
by lumost 3y ago
New deal was about 90 years ago. It doesn’t seem believable that the ideal form of the US government was a situation from 90 years ago. Hopefully this is a case where the ruling would be defacto recognition of the modern agency as a delegat

More ›