valine

28 ms·

331.
by valine 3y ago
That’s a lot of words, but at the end of the day the NHTSA allows FSD beta on the roads. Someday in the distant future Tesla will likely use the data they’ve collected to make statistical inferences to regulators about the safety of the sys
332.
by valine 3y ago
Those concerns are entirely unwarranted, boarding on hysteria. FSD has an adequate real world safety record.
333.
by valine 3y ago
Nothing you said is wrong, I just really, truly don’t care. FSD makes fewer mistakes on routes I drive than it did a year ago, so whatever they’re doing seems to be working. Complaining that people like me shouldn’t be allowed to purchase s
334.
by valine 3y ago
It kind of works, that’s the point. I pay for an FSD subscription because I get enjoyment from my car driving me around. Am I being scammed?
335.
by valine 3y ago
Tesla gets a bad rap because of their notorious missed deadlines. They continue to chip away at the problem though, and have a large number of people paying them to use their beta software. The fact that they’ve figured out how to make inco
336.
by valine 3y ago
Better quantization will have a direct and meaningful impact for everyone running local LLMs. The technique has already been applied to both Vicuna and the multimodal LLaMA variant LLaVA. https://github.com/mit-han-lab/
337.
Activation-Aware Weight Quantization for LLM Compression Outperforms GPTQ(arxiv.org)
9 points by valine 3y ago|2 comments
338.
by valine 3y ago
This watermark relies on subtle grammatical variations. Passing it through any model is going to wipe out the distribution. The number of open LLMs is exploding, and the most popular ones are fine tuned by small groups / individuals. N
339.
by valine 3y ago
GPT-4 wasn’t nerfed. https://news.ycombinator.com/item?id=36155267
340.
by valine 3y ago
If your goal is to track novice users there are much easier methods. You could insert invisible unicode characters for example.
341.
by valine 3y ago
What’s to stop someone from taking the watermarked output and randomizing the distribution by feeding it through their latest LLaMA variant? These watermarks will only be useful for catching novice LLM users.
342.
by valine 3y ago
VR hardware is chock full of compromises right now. VR’s final form factor will probably look more like goggles or glasses.
343.
by valine 3y ago
Location: Minneapolis, Minnesota Remote: Yes Willing to relocate: Yes Technologies: Machine Learning, C++, Python, PyTorch, CUDA Résumé/CV: https://zpath.org/cv.pdf Email: lukas@valine.io I’m a software engineer with o
344.
by valine 3y ago
Its hard to make other use cases work when text is barely readable. The resolution just isn't there for productivity apps. I'd be surprised if Apple doesn't lean into that angle with their new headset. Pixel density on phones
345.
by valine 3y ago
That would work for inference, but for efficient training you’d also want you training set to fit in memory.
346.
by valine 3y ago
There were rumors floating around that GPT-4 was going to be a 100 trillion parameter model. Those rumors seemed ridiculous in hindsight, but this announcement makes me rethink how ridiculous it really was. 100 Terabytes of GPU memory is ex
347.
by valine 3y ago
It's not clear at all that we have an avenue to super intelligence. I think the most likely outcome is that we hit a local maximum with our current architectures and end up with helpful assistants similar in capability to George Lucas&
348.
by valine 3y ago
I have yet to see a solution for “AI safety” that doesn’t involve ceding control of our most powerful models to a small handful of corporations. It’s hard to take these safety concerns seriously when the organizations blowing the whistle ar
349.
by valine 3y ago
I’m also fairly certain an FFT of the audio would instantly uncover the repeating patterns. The repeating tones would show up as taller spikes.
350.
by valine 3y ago
Does anyone know of a PCIE 5 SSD designed for sustained read/write? Most of these new drives are meant for short bursts of data transfer, I can’t imagine water cooling would be necessary for most drives.
351.
by valine 3y ago
At the risk of anthropomorphizing, GPT-4 is smarter. It writes better code, hallucinates less, and is just generally more helpful.
352.
by valine 3y ago
Vertical integration has many benefits. It’s the same reason Apple started making their own chips and ditched intel. The SOCs will be cheaper and better optimized for Tesla’s use case. Every Nvidia GPU I’m aware of ships with both CUDA core
353.
by valine 3y ago
Probably because the world is transitioning to EVs and they have a 5ish year head start over most other automakers. There’s a long list of technologies Tesla has already shipped that other car makers are just beginning to explore. - Structu
354.
by valine 3y ago
Why are you loitering beside other cars? This is a good safety practice regardless of the car’s make and model.
355.
by valine 3y ago
Fine tuning is moderately intense. Going from base LLaMA to Vicuna 13B costs about $300.
356.
by valine 3y ago
In a perfect world you give the LLM a python interpreter and it does the rest. Realistically, with the LLMs we have today, the right approach is probably to curate a set of APIs it's allowed to interact with. Some basic file system acc
357.
by valine 3y ago
We need something like this on Linux, maybe powered by Vicuna. I’m not sure if the current batch of LLaMA variants is coherent enough to work as a digital assistant, but my gut feeling is that a little fine tuning on tool use might be all t
358.
by valine 3y ago
Useful LLMs are probably still a tad to large to run locally on device. Loading and unloading the model takes several seconds, so ideally the model would just sit in RAM at all times. In my experience playing with local LLaMA variants, Chat
359.
by valine 3y ago
Multiple choice tests are easier than fill in the blank precisely because it's easier to recognize when something is correct than it is to regurgitate a fact from thin air. You don't have to be an expert to recognize when ChatGPT
360.
by valine 3y ago
Why is no one mentioning the hardware requirements to run each? The semantic version runs with 2KB of RAM and takes a couple of microseconds. The LLM needs at least 30GB of RAM and maybe 5 seconds of processing time.

More ›