kiratp

9 ms·

91.
by kiratp 3y ago
Because we don’t have a clear understanding of how the things work yet.
92.
by kiratp 3y ago
It’s not reasonably possible (currently?) to get the same performance from a 7 billion parameter model as a 175 billion parameter model with just an additional 6000 lines of finetuning data.
93.
by kiratp 3y ago
Current LLMs can’t do “random”. There are 2 sources of randomness: 1) the random seed during inference 2) the non-determinism of GPu execution (caused due to performance optimizations) This is one of those things that humans do trivially bu
94.
by kiratp 3y ago
Given that Apple has built PyTorch for Metal, I would assume they handle unified memory properly in there too… Same for Tensorflow and Jax on Metal.
95.
by kiratp 3y ago
I wish there was more information as to how this differs from or improves on Jax. Flax+Jax+OpenXLA seems to finally be building some momentum so when a big player launches yet another competitor, the justification for it would be a good thi
96.
by kiratp 3y ago
Because the boomers vote and the younger generations do not.
97.
by kiratp 3y ago
But the likes of GM do? https://www.atlantamagazine.com/great-reads/no-accident-insi... > it meant GM had produced a part that failed to meet not only its own internal minimums for torque but also basic federal safe
98.
by kiratp 3y ago
Those same people will microwave a meal for dinner and then make breakfast eggs on an anon-stick Teflon pan the next morning.
99.
by kiratp 3y ago
> Every Falcon (or Starship) carrying KuiperSats is _not_ carrying Starlink satellites. SpaceX uses Starlink launched as a way to tune and optimize the Falcon launch platform. Launching KuiperSats potentially hurts the SpaceX launch vehi
100.
by kiratp 3y ago
> Kuiper particularly looks like a long shot, because it's directly fighting against SpaceX's biggest strength: large-number launch economics. Of all the ways to try to get a foothold in the new space industry, the Kuiper appro
101.
by kiratp 3y ago
There are so many billions of dollar being spent on this hardware that everyone other than Nvidia is doing everything they can to make competition happen. Eg: https://www.intel.com/content/www/us/en/deve
102.
by kiratp 3y ago
You can't make this optional because they will be "taught" how to do it by the likes of Epic and Facebook. If its available people will be exploited. Has already happened: https://grahamcluley.com/android-secu
103.
by kiratp 3y ago
Software engineering is first and foremost about making good engineering choices. We are proving time and again as an industry that humans CANNOT write mistake-free code. From just this week: https://news.ycombinator.com/ite
104.
by kiratp 3y ago
Written in C++ is not a positive in my book. New things created this decade in unsafe languages (where safe options would have worked fine) should be frowned up and criticized as bad engineering.
105.
by kiratp 3y ago
We don't get by fine on x86. That is the whole point here. My non-technical family are able to make their iOS devices "just work" with 0 support.
106.
by kiratp 3y ago
This thing is basically closed source and closed license. The license explicitly blocks “Hosting Use” and does not even differentiate between “model as an API” and in-application use. I hope the community rejects this kind of BS.
107.
by kiratp 3y ago
Two 3090s don’t have 96GB of VRAM
108.
by kiratp 3y ago
> One way I could do this is by completely implementing its wire-format API of MongoDB as a proxy server to my derived MongoDB, and offer that up as my SAAS This is what Cosmos DB for Mongo DB is (wire format endpoint that talks to Msft’
109.
by kiratp 3y ago
India’s obsession with SMS codes and UPI in particular sucks as a tourist/visiting businessperson. Nothing works with a non-Indian number. You can’t get an Indian SIM without an Aadhar number outside an airport and every time I’ve flow
110.
by kiratp 3y ago
> Using that dominance in one market to crush competitors in an unrelated market is generally a big no-no. Phone lines have always been part of a functional office. Not sure how you can argue that communications (voice, video, screen sha
111.
by kiratp 3y ago
> it can't be that difficult to port whatever code to whatever platform if 100s of millions are in question, or is it? 300M hires a lot of engineers. Maybe the fact that they are buying instead of building tells you how hard it is t
112.
by kiratp 3y ago
Here is another perspective. The web is already FULL of bad medical information. Older folks treat it as gospel already. All this technology does is make both good and bad information more accessible. More importantly, there is NOTHING anyo
113.
by kiratp 3y ago
I love Huggingface. I worry they will be the next Docker. What is the moat? "We will run your inference" can't be the answer.
114.
by kiratp 3y ago
> We are only missing faster local development. Developer time is probably your most expensive resource. Saving $8/mo at the cost of dev time seems… suboptimal.
115.
by kiratp 3y ago
> Rust doesn't have a "I know what I'm doing, let me be" switch. Every piece of code you write is security-sensitive. IMO It is our professional responsibility to our users that we treat it that way. For a vast majori
116.
by kiratp 3y ago
Rust is not a GC language. It can and does achieve the same perf as any modem C++ with the latest flavor of the C++ safety guardrails. The set of reasons to start a new project in C/C++ are few and the list is shrinking by the day.
117.
by kiratp 3y ago
OpenAI APIs have pretty much as clear a contract as you can get with a third party. > Starting on March 1, 2023, we are making two changes to our data usage and retention policies: > OpenAI will not use data submitted by customers via
118.
by kiratp 3y ago
> Starting on March 1, 2023, we are making two changes to our data usage and retention policies: > OpenAI will not use data submitted by customers via our API to train or improve our models, unless you explicitly decide to share your
119.
by kiratp 3y ago
IMO don’t try to be the one stop shop to host models. There are too many players with all sorts of advancements (eg: stopping grammar, continuous batching, novel quantization etc.) and you won’t be able to keep up. There is a ton of boilerp
120.
by kiratp 3y ago
This would be super useful if it supported local/in-K8-cluster models. Most production use cases probably need some sort of fall back of small -> medium -> large -> GPT4. Given the costs + low quota limits, I would be surprise

More ›