gaogao

14 ms·

91.
by gaogao 3y ago
A lot of top ML/AI talent has already bailed too, so some of it is probably them trying to keep open research closer to SOTA.
92.
by gaogao 3y ago
Their consent order with the FTC also contains a prohibition against privacy misrepresentations, so it would probably get audited during their biennial assessments. For some unethical company that doesn't get regularly audited, yeah th
93.
by gaogao 3y ago
Its metrics on HumanEval seem not particularly good (26.89 Pass@1 for it vs. 61.64 for PanGu-Coder2 15b). Is it targeting a very specific latency for responses? I'd think a 15b quantization should run fast enough for most use cases? Ev
94.
by gaogao 3y ago
Huh, kind of weird that they don't have the chat-tuned Llama 2 in the comparison mix.
95.
by gaogao 3y ago
I have this pet proposal that LLMs would be pretty nice to help fill out WikiData https://friend.computer/jekyll/update/2023/04/30/wikidata-ll... , as the technique of getting LLMs to write queries,
96.
by gaogao 3y ago
On the cyberpunk side, the LLM demoscene is also incredibly fascinating. We folks beating out the megacorps on model hacks like NKT and hecking crazy model names like TheBloke/WizardLM-Uncensored-SuperCOT-StoryTelling-30B-SuperHOT-8K-G
97.
by gaogao 3y ago
We have both itty bitty calculators and supercomputers, so pretty feasible to have both edge AIs and central ones.
98.
by gaogao 3y ago
The unicorn example from MSFT's Sparks of AGI paper, https://arxiv.org/abs/2303.12712 , probably counts as a novel problem. As another example, ImageBlind, https://ai.facebook.com/blog/imagebin
99.
by gaogao 3y ago
LLMs and Semantic Web work reasonably well in concert https://friend.computer/jekyll/update/2023/04/30/wikidata-ll...
100.
by gaogao 3y ago
They want that juicy RLHF training data
101.
by gaogao 3y ago
Bard doesn't even best 3.5 save for having knowledge base lookups
102.
by gaogao 3y ago
I did actually run into their problem on my gpt-4 prompt, but it was reasonably correctable.
103.
by gaogao 3y ago
That specific case was it using too much common sense (which was apparently also a problem with IQ tests at one point) Me: If it rains today, I will drive home. I did not drive home. Did it rain today? Answer using only logic. ChatGPT: Usin
104.
by gaogao 3y ago
> What about simply asking the LLM to craft its own search prompt to the DB given the user input, rather than returning articles that semantically match the query the closest? I played with that approach in this post - https:/&#x2F
105.
by gaogao 3y ago
WikiData also works pretty well too - https://friend.computer/jekyll/update/2023/04/30/wikidata-ll... I've been thinking that you probably can do something better than just embedding search on
106.
by gaogao 3y ago
Hmm? You can get the raw data and data for tables too. I've done a full dump before for offline analysis. The API for getting it slightly more in native format than just CSV and Markdown works fine too. There's a couple of Notion
107.
by gaogao 3y ago
Likely not, but it depends on the language around anonymizing that the bill uses
108.
by gaogao 4y ago
(2022)
109.
by gaogao 4y ago
In a few of the examples given, WikiData would also be searchable enough to find the relevant articles to pull in. I've actually been really curious if you could use recent advancements in ML to augment the filling out of WikiData, sin
110.
by gaogao 4y ago
Still getting data until you need to boom it
111.
by gaogao 4y ago
Where do you see ChromeOS and Google Office as falling flat in that regard? Desktops are pretty smaller in market size compared to laptops these days.
112.
by gaogao 4y ago
In terms of the privacy policy, how are y'all making sure queries containing PII get anonymized?
113.
by gaogao 4y ago
It's fairly similar to a ROM patch in the video game space, which has mostly stood the test of time.
114.
by gaogao 4y ago
Yeah, my sense is that OpenAI moat is primarily just through the RLHF dataset right now. Most of the other things – foundational models and datasets, embeddings (roughly the plugins announcement today) – have generally already been commodit
115.
by gaogao 4y ago
It has data as of yesterday, ingested through Knowledge Graph I think (could potentially be something more raw too). For example, it can answer stock price questions from yesterday.
116.
by gaogao 4y ago
You could do attested code in an enclave, which though vulnerable to certain side channels, is probably more robust than the standard case.
117.
by gaogao 4y ago
Has anyone tried this yet on the 65B version? I'm curious if it knows how to rhyme and other emergent behavior, as alpace-7B does not.
118.
by gaogao 4y ago
> but build me a consistent world which spoons out the world-building _and remembers it_ Yeah, one potential way to do this is to add input / output to a semantic graph store a la WikiData, in addition to the more bog standard text
119.
by gaogao 4y ago
Wagering a guess, are you exposed to way more spoken Japanese media through anime than you are spoken Spanish content?
120.
by gaogao 4y ago
Meta has released the model weights for OPT-175B, which is used in the paper. There's also a lot of full release LLMs from other labs on the way as well.

More ›