HNHacker News
TopNewBestAskShowJobs

letitgo12345

416 karma · joined December 26, 2014

submissionscomments
letitgo12345··on What it feels like to work in AI right now
Issue is the data moat OAI is building. They'll have hundreds of millions of high quality user interactions with ChatGPT they can use to finetune their models. What will anyone else including Google have?
letitgo12345··on What it feels like to work in AI right now
People really don't seem to understand just how far ahead OAI is in its reasoning abilities (https://crfm.stanford.edu/helm/v0.2.2/?group=reasoning)

Maybe PaLM is near there (it's not evaluated on that page) but nothing else even comes close at all

letitgo12345··on What it feels like to work in AI right now
Hope so. Don't want a single corporate entity (OAI/MS) dominating the entire economy. This sector desperately needs competition
letitgo12345··on What it feels like to work in AI right now
"I'm looking for X. Can you give me a few options with the benefits and drawbacks of each"

or

"I'm looking for X. Can you ask me a few questions and give me a choice of products based on that"

or

"I'm looking for a product that does X, Y, Z. Can you find such a product for me"

or

"Does this product do Y/is compatible with Z/is an appropriate give for person P"

I'm not saying this will happen. I'm saying it's a risk that Amazon should take seriously.

letitgo12345··on What it feels like to work in AI right now
From what I heard, they don't have scaling capabilities to deploy the bigger models right now which they're working on and that's the biggest issue.
letitgo12345··on What it feels like to work in AI right now
They have been collecting human feedback data for 2 years + probably have a lot of data from Copilot + are training with large context models + have invested a ridiculous amount in curating pretraining data -- the kind of stuff that won't get you a ton of pubs (so you won't see Google researchers having focused on it a lot) but apparently turns out to be super important for a good LLM
letitgo12345··on What it feels like to work in AI right now
Imagine Walmart launching a ChatGPT interfaced bot for shopping that customers take a liking to. Walmart starts acquiring both new customers as well as high quality data they can use for RLHF for shopping. Eventually Walmart's data moat becomes so big, that Amazon retail cannot catch up and customers start leaving Amazon.

For AWS, if MS starts giving discounts for OAI model usage to regular Azure customers, that's gonna be a strong incentive to switch

For Apple, A Windows integrated with GPT tech may become a tough beast to beat.

letitgo12345··on GPT-4 performs significantly worse on coding problems not in its training data
I would prefer to say that we've seen a glimpse of what a future world with a perfect reasoner will be like
letitgo12345··on GPT-4 performs significantly worse on coding problems not in its training data
It's definitely not just copying verbatim. If you mean it's emulating the reasoning pattern it sees in the training data well...don't humans do that as well to get answers to novel problems?
letitgo12345··on GPT-4 performs significantly worse on coding problems not in its training data
I guess the meaning of our existence was always handing the universe over to the AI overlords lol
letitgo12345··on GPT-4 performs significantly worse on coding problems not in its training data
Will take a bit of time before AI can consistently beat us on coding/proofs but the raw ingredients imo are there. As someone who was skeptical of AGI via just scaling things up even after GPT-3, what convinced me was the chain of thought prompting paper. That shows the LLM can pick up on abstract thought and reasoning patterns that humans use. Only a matter of time before it picks up on all of our reasoning patterns (or maybe it already has and is just waiting to be prompted properly...), is hooked up to a good memory system so it's not limited by the context window and then we can watch it go brrrr

It can still make stupid mistakes in reasoning but I don't think that's fundamentally unsolvable in the current paradigm

letitgo12345··on ChatGPT Plugins
They sank it in "customer obsessed research" lol
letitgo12345··on Sparks of Artificial General Intelligence: Early Experiments with GPT-4
Where are you getting the 200 number from?
letitgo12345··on Bard is much worse at puzzle solving than ChatGPT
Even the best Google models seem to be lagging for reasoning tasks vs OpenAI ones at the moment - see the graphs at https://github.com/suzgunmirac/BIG-Bench-Hard
letitgo12345··on OpenAI to discontinue support for the Codex API
They're paid so presumably safer at least
letitgo12345··on OpenAI to discontinue support for the Codex API
Ughh and they don't offer logprobs for the chat interface...
letitgo12345··on OpenAI to discontinue support for the Codex API
It's 10 cents/1K tokens - far cry from the 2 cents/1K tokens that equivalent size models OpenAI was offering
letitgo12345··on Amazon’s big dreams for Alexa fall short
> It is stunning that Amazon wasn't the forerunner in using ChatGPT/LLM-based tech for Alexa

I'm a current scientist at Amazon. Amazon has little ability to invest in longer term research. While you have scope to pursue projects without immediate product application to some extent (though far less than you would at Alphabet/Meta/Microsoft), that is less valued than at the other Big Tech companies. There's also a far more bureaucratic hurdles to jump through to publish at Amazon. The lack of care for non-product research shows at every instance. Amazon also has extremely low -- or I should say non-existent standards for hiring scientists. If you're putting a person with a decent and non-existent pub record on the same level, that is not going to yield great results.

Luckily for Amazon, their bread and butter -- Ads via retail and AWS -- are not immediately threatened by LLMs.

Honestly if they had taken the billions they burnt on Alexa and put it in research they would be far far far better off today. But try explaining that to the bean counter mentality of executives that Bezos cultivated...

letitgo12345··on TensorFlow for Python is dying?
Theano preceded Tensorflow by 8 years. So I don't think it was all manual coding before TF. Torch (in Lua) was also around much before TF
letitgo12345··on Copilot Internals
You would be taking snippets of code (that are potentially unparseable), concating them together and putting it in the prompt. The issue is that it would be kind of prompt the model has never seen before in the training data. Maybe it would work with some clever 0-shot prompt. But if you look at the fill in the middle paper from OpenAI for example, they specifically pretrain the model with that kind of data to make it work.

The live data is gonna be useful though ya. Is Copilot allowed to use it though under ToS?

letitgo12345··on Copilot Internals
Not sure how easy it would be to make work. Code edit data is not that prevalent. The best I can think of is looking at github commit changes. That's one place where Repl.it has a big advantage as it has live editing data from its users
letitgo12345··on Discovering faster matrix multiplication algorithms with reinforcement learning
I think they find a N^2.778 complexity algorithm (obtained by applying their 4x4 discovered algorithm recursively) so this is not correct.
letitgo12345··on Facebook shuts out NYU academics’ research on political ads
The realistic future is Chinese social media companies (and by proxy Chinese govt) ruling the internet instead of Facebook and Twitter :)
letitgo12345··on Our brutal science system almost cost us a pioneer of mRNA vaccines
Flip side -- the pioneering research may not have been possible without govt funding of science through a grant based system. Imo the right conclusion here is to increase govt funding massively. Unclear to me how to not do it based on a grant system. Even in the industry you need to do project proposals to keep your team funded.
letitgo12345··on Cryo–electron microscopy breaks the atomic resolution barrier
It's a bit trickier than that as the flash freezing process is on the scale of milliseconds iirc which can affect what ensemble of folds are seen via cryo-em
letitgo12345··on Madhya Pradesh man cycles 105km to ferry son to Class X exam centre
Maybe the reason some Asians are successful given the opportunities is because such stories are thought of as uplifting instead of lamenting against the system. The system is bad but what's wrong with celebrating and finding inspiration from people who beat it?
letitgo12345··on MIT moves all classes online for the rest of the semester
MIT has an exception for homeless students.
letitgo12345··on Coronavirus Spike Protein Binder Design
I doubt it's easy to isolate
letitgo12345··on Google employee diagnosed with coronavirus
Even apart from undercounting of case #s, we know that the # of severe cases is 5% upper bound which is much more likely to be accurate and the deaths are a fraction of that so I would bet on the number being on the lower end than the higher end. Even a 1-2% mortality rate is fairly substantial tho -- 5-10x as deadly as flu.
letitgo12345··on SpaceX tests black satellite to reduce ‘megaconstellation’ threat to astronomy
My impression is that a lot of interesting "leads" are generated by ground based telescopes and only then examined in more detailed by space based telescopes. If you kill off the generating mechanism, then effectiveness of space telescopes could also decline a lot.
← PreviousPage 3 of 7Next →