HNHacker News
TopNewBestAskShowJobs

PheonixPharts

4,161 karma · joined April 7, 2022

submissionscomments
PheonixPharts··on Ice-free summers in Arctic possible within next decade, scientists say
You can always find quotes from people who are off in their estimates, but even a glance at the data [0] should convince you that things are not going well and that all of these predictions were at the very least directionaly correct.

Genuine question, have you never looked at the actual data behind those claims, or have you seen it and still don't find it convincing? My practice has been that whenever I see a sensational headline (or even academic article), I go dig up the underlying data and see if it's convincing. While undoubtedly media reporting is sensationalist (both direction depending on the news), it didn't take long plotting a wide range of different data sets to convince me that things are not looking good.

And that was ~10 years ago. I've personally been shocked how rapidly the climate has measurably and visibly impacted human life since then.

0. https://www.arcticdeathspiral.org/

PheonixPharts··on Ice-free summers in Arctic possible within next decade, scientists say
"20 years ago they said these events were 30 years away, and now they're saying they're just 10 years away, can't these climate scientists ever get it right!"
PheonixPharts··on The Black-Scholes/Merton equation [video]
The underlying assumption Black-Scholes makes, that stock price movements can be modeled by a log-normal distribution, is known to be false. However not since the 1980s has this lead to the ability to make money of the model itself being imperfect.

The true distribution of the market beliefs in future stock prices can be understood by empirically studying the volatility smile [0]. That is, because investors know Black-Scholes is not a perfect mathematical model of real world stock behavior, every strike price has a different implied volatility. By looking at these different IVs you can get a sense of what the market believes are the true probabilities of "long tail" events.

In theory, the opportunities you have to make money should be cases where you believe the market has mispriced risk. In my amateur experience, I have found that virtually every time you think the market has mispriced some extreme event, when you look at the volatility smile, you realize you are mistaken.

0. https://en.wikipedia.org/wiki/Volatility_smile

PheonixPharts··on The Black-Scholes/Merton equation [video]
Adding to this, it's very worthwhile exercise for any curious programmer to work out the IV of a stock based on options pricing and compare it to other measure of volatility (for example historic volatility, or even your own beliefs about volatility based on what you think future returns might be).

Black-Scholes/Merton makes a lot more sense once you work it all out yourself in code.

I'd actually suggest doing this through modeling the underlying geometric Brownian motion and ensuring that your simulated results match up to the analytic formula.

PheonixPharts··on Money bubble
> Customer support team of 10 is replaced with chatbot and 2 people.

My experience, having grown up when all customer service reps were people, is exactly what I stated above: this is just giving up on customer service.

Anyone who has ever called automated support knows this. When you reduce 10 people to 2 people and some chatbots now you simply have to wait 5x as long for customer support.

I worked at startup a few years back that refused to scale customer support so they could be forced to "automate" the process. The result? Customers got completely screwed over, but those customers weren't investors so who cares.

I can't recall a single time in my life were automated customer support solved my problem, it just kept me busy so that the 5x wait doesn't seem as long since I'm trying to navigate the labyrinth of a customer support decision tree to get my problem solved.

PheonixPharts··on Money bubble
Anyone who works in AI, especially very closely with models, will tell you that it's not capable of really replacing any jobs yet.

All these jobs being "replaced by AI" are simply being eliminated with the consequences of them being eliminated ignored. Customer service jobs aren't being replaced by AI, companies, like Klarna, are just giving up on customer service and using AI to increase their perceived value rather than reducing it.

PheonixPharts··on Money bubble
> Things have been too good for too long in InvestorWorld

That's not by accident and it's been at the expense of non-investor world for a long time.

What we're seeing is finance capitalism suck surplus value from every piece of the Earth it can. We're still burning more fossil fuels than ever before [0] despite the now visible risk of climate catastrophe. I believe we're likely to see investors continue to become irrationally wealthy while increasing larger and larger pools of people a driven into ever increasing situations of hardship.

I see no signs of the madness stopping until both human and planetary resources start to buckle under the pressure and refuse to give yields they once did. The article repeatedly mentions crypto as though it were an obvious bubble, but bitcoin is near record highs and even Sam Altman's bizarre world coin is at extreme record highs, COIN is up 200% in the past year.

The bubble won't "burst", it's just that increasingly less people will be invited in.

0. https://ourworldindata.org/fossil-fuels

PheonixPharts··on Apple to wind down electric car effort after decadelong odyssey
I do a lot of open source LLM research/dev work on a Mac Studio. While it doesn't quite compete in terms of speed with a GPU for standard transformers models, I can run pretty huge models locally. When I'm working with llama.ccp, the speed and model size I can run is very impressive.

I can also run Stable Diffusion XL in reasonable time frames on an iPad Pro. The current gen Macbook Pro can perform almost as well as the Mac Studio with an M2 Ultra, only the M3 Max has about 1/2 the bus speed (though still wild that you can run good sized local LLMs on a laptop).

If local generative AI becomes a major part of computing in the future, Apple has a huge advantage over the other players out there. This was obvious the second I started working on my Mac Studio. I have spent plenty of time using a traditional GPU setup for LLM work, and yes it is faster, but the complexity of getting things running is way beyond the average user's ability. Not having to fight with cuda ever is amazing, and so far everything else has 'just worked' as is typical of Apple.

If Apple has a team of talented people working to get gen AI performance tuned specifically to their hardward, I suspect we'll see some very competitive offerings in this space.

PheonixPharts··on Apple to wind down electric car effort after decadelong odyssey
> They literally wrote the book on this

I didn't realize Apple literally wrote a book on the subject. Sounds interesting, what's the title?

PheonixPharts··on Why jalapeño peppers are less spicy (2023)
I think you're misunderstanding what I'm saying. I enjoy plenty of non-spicy foods, and don't need every meal to be spicy, nor to I think that the only flavorful meals are spicy meals.

My point is exclusively with chili peppers, as the spice increases so does the flavor. I didn't grow up eating spicy food, and my experience has been that as my spice tolerance has grown overtime I have found access to increasingly wonderful flavors that are simply not present at lower spice levels.

PheonixPharts··on Why jalapeño peppers are less spicy (2023)
> I prefer more flavorful chiles, with less spice.

As someone with a fairly high spice tolerance, I've found a strong correlation between heat and flavor. I think anyone coming from a culture that eats spicy food will agree.

I had a batch of habaneros I grew last summer that were both the most spicy habaneros I've had, and by far the most flavorful. I've also had some very tasty ghost peppers (when used in moderation). Outside of peppers that are inherently not spicy, such as pablanos, heat and flavor seem to coincide.

All of the most flavorful chili dishes I've had in my life have all been quite spicy, and I can't recall any mild chili dishes I've had with nearly the same punch.

The only time I've seen heat and flavor not connected is in terrible hot sauces which aim to be shockingly spices that make use of capsaicin extract.

PheonixPharts··on Microsoft strikes deal with Mistral in push beyond OpenAI
You're really generating $3000 per month from ChatGPT? Can you give a hint about what you've built that generates this kind of ROI?

I have only seen people making money in AI by selling AI products/promises to other people who are losing money. The practical uses of these tools still seem to be largely untapped outside of as enhanced search engines. They're great at that, but that does not have a return on value that is in proportion to current investment in this space.

PheonixPharts··on Microsoft strikes deal with Mistral in push beyond OpenAI
I've stopped believing in rational markets a long time ago. For the people behind this funding madness:

> to skew markets, skew the financials of big tech and create a bubble in the space.

Are the intended consequences of this. The people behind the money in AI, just like the people behind the money in crypto, don't care if there's a reality to all of this, they just care if they can make a lot of money while the music is playing.

I really thought 2022 was going to be the beginning of tech returning to reality, but naively didn't understand that this would entail a lot of people with a lot of money losing money, and that's not going to happen.

As with all bubbles, the interesting thing isn't pointing out there's a bubble, we've been living in many bubbles for decades now. The interesting thing is pointing out what will make it pop. So long as globally money keeps pouring into US markets we'll see this continue.

On the plus side, at least LLMs are a lot of fun to work and play with!

PheonixPharts··on Every model learned by gradient descent is approximately a kernel machine (2020)
You must have a pretty impressive resume if you don’t consider François a researcher! The wikipedia would disagree with you [0], as would anyone that has had any interaction with him on the subject.

0. https://en.wikipedia.org/wiki/Fran%C3%A7ois_Chollet

PheonixPharts··on Every model learned by gradient descent is approximately a kernel machine (2020)
François Chollet, author of Keras, lays it out pretty clearly that LLMs are just memorizing [0] (there’s more links if you follow that thread).

I think LLMs are incredible, and spend most of my days working with them closely, but they are not nearly as close to “AGI” as people think primarily due to their inability to really generalize.

At the end of the day LLMs aren’t that different than old school n-gram Markov chains, except rather than working on n-grams, they’re working in a (very sophisticated) latent space. Their power is really these incredible latent languages spaces we’re still just starting to understand.

In all my years of tech the “AI” space is the most curious hype-bubble since the things people expect to happen are entirely out line with what is possible, while at the same time the potential of these models is still, imho, underexplored and largely ignored by the vast majority of people attempting to build things with them.

99% of the people I know working in this space are just calling APIs and trying to do some variant of code generations, where a small minority of people are really trying to figure out what’s going on in these models and what can be done with them successfully.

0. https://twitter.com/fchollet/status/1755250582334709970

PheonixPharts··on The reasons people are sending back their Apple Vision Pro headsets
You don't think the even simpler answer is that VR is really not a solved problem yet for everyday users?

I really enjoy my psvr2, but I'm never going to have that headset on for more than an hour a day every few days. I thought the Quest was a complete game changer for the first month, then put it away never to be used again.

It looks like both Apple and Meta have moved the needle on the overall experience, but it seems that there is still a long way to go to get VR to the, perhaps impossible, point where it becomes as ubiquitous as mobile phones. VR remains a non-essential novelty and likely might forever remain such.

PheonixPharts··on Machine Assisted Proof [video]
It's not about not trusting it or "staying in your lane", LLM results for domain specific work are terrible, and additionally useless if you're writing fairly novel code (which most ML practitioners/Compiler programmers are doing if they're doing interesting work).

The usefulness of LLMs for writing code is strongly correlated with how "Stack Overflowable" your work is.

PheonixPharts··on Ask HN: What are some actual use cases of AI Agents right now?
This thread is an object lesson in the point I'm making: people have forgotten everything we've learned about making ML based products.

Parent comment doesn't understand the concept of expectation, and this comment is apparently unfamiliar with the fact that SotA for digit recognition [0] has been much higher than 90% even in the 90s. 90% accuracy for digit recognition is what you get if you use logistic regression as your model.

My point was that numbers that look good in terms of research often aren't close to go enough for the real world. It's not that 90% works for zip codes, it's that in the 90s accuracy was closer to 99%. You have validated my point rather than rejected it.

0. https://en.wikipedia.org/wiki/MNIST_database

PheonixPharts··on Ask HN: What are some actual use cases of AI Agents right now?
> I'd love to know what are some of the hidden challenges to making a useful product with agents?

One thing that is still confusing to me, is that we've been building products with machine learning pretty heavily for a decade now and somehow abandoned all that we have learned about the process now that we're building "AI".

The biggest thing any ML practitioner realizes when they step out of a research setting is that for most tasks accuracy has to be very high for it be productizable.

You can do handwritten digit recognition with 90% accuracy? Sounds pretty good, but if you need to turn that into recognizing a 12 digit account number you now have a 70% chance of getting at least one digit incorrect. This means a product worthy digit classifier needs to be much higher accuracy.

Go look at some of the LLM benchmarks out there, even in these happy cases it's rare to see any LLM getting above 90%. Then consider you want to chain these calls together to create proper agent based workflows. Even with 90% accuracy in each task, chain 3 of these together and you're down to 0.9 x 0.9 x 0.9 = 0.73, 73% accuracy.

This is by far this biggest obstacle towards seeing more useful products built with agents. There are cases where lower accuracy results are acceptable, but most people don't even consider this before embarking on their journey to build an AI product/agent.

PheonixPharts··on Hi everyone yes, I left OpenAI yesterday
I think money is the least of the reasons why those founders wouldn't attract him.

To attract someone at Karpathy's level you would need a project that is both wildly challenging (and yet not the typical startup "challenging" because it's a poorly thought out idea) and requires the kind of resources (compute, data, human brains in vats, etc) that would make your place look far more interesting than OpenAI.

But, hardest of all, you would need startup founders that could tame their egos enough to let someone like Karpathy shine. I haven't talked to a Bay area startup founder in a while who wouldn't completely squander that kind of talent by attempting to force their own half-baked ideas on him, and then try to force him out months later when he couldn't ship those poorly thought out products citing lack of "leadership".

PheonixPharts··on Bioluminescent petunias now available for U.S. market
Just wait until you learn about tulips! I never understood tulip mania until my wife started planting them.

Tulips are annuals, so they only are expected to grow one season. You often order the bulbs in summer and they're delivered in fall, and then planted in early winter before the ground gets too cold. They're often less than a dollar per bulb, but if you've ever seen a nice arrangement, hundreds of tulips is not all that much.

The first year my wife told me she was spending ~$300 on tulips I thought she was crazy. That spring when I saw what they looked like I okayed doubling the order for the next year. It's really a delight to see spring announced by your yard exploding in beautiful flowers.

PheonixPharts··on Show HN: Faster LLM evaluation with Bayesian optimization
"Evaluation" has a pretty standard meaning in the LLM community the same way that "unit test" does in software. Evaluations are suites of challenges presented to an LLM to evaluate how well it does as a form of bench-marking.

Nobody would chime in on an article on "faster unit testing in software with..." and complain that it's not clear because "is it a history unit? a science unit? what kind of tests are those students taking!?", so I find it odd that on HN people often complain about something similar for a very popular niche in this community.

If you're interested in LLMs, the term "evaluation" should be very familiar, and if you're not interested in LLMs then this post likely isn't for you.

PheonixPharts··on Neural network training makes beautiful fractals
I find this result absolutely fascinating, and is exactly the type of research into neural networks we should be expanding.

We've rapidly engineered our way to some very impressive models this past decade, and yet gap in our real understanding of what's going on has widened. There's a large list of very basic questions about LLMs that we haven't answered (or in some cases, really asked). This is not a failing of people researching in this area, it's only that things move so quickly there's not enough time to ponder things like this.

At the same time, the result, unless I'm really misunderstanding, gives me the impression that anything other than grid search hyper parameter optimization is a fools errand. This would give credence to the notion that hyper parameter tuning really is akin to just re-rolling a character sheet until you get one that is over powered.

PheonixPharts··on "Accelerationism" is an overdue corrective to years of gloom
"Years of gloom"?

For my entire career in tech, anything that isn't the most extreme of optimistic view points has been relegated to unnecessary and destructive pessimism.

Tech can't stand "bad thoughts". We had a very necessary correction in this industry but people couldn't handle it and just completely lost their minds on AI.

I say this as someone who works in AI very close to the metal (i.e. not just calling an API all day). While these tools are very impressive, when I talk to people outside of my area it's like they've all taken obscene quantities of some new stimulant and can't even connect with reality anymore.

Even the AI "doomers" are, by my view, extreme optimistic because they don't see, or maybe just don't want to see, how much all of the public discourse on this subject is largely smoke and mirrors to drive yet another tech bubble.

Personally I think this is just mass hysteria created by the increasing awareness of the fundamentally unsustainable nature of industrial society. Reading this article is like listening to some hallucinating maniac on the streets of SF scream at you about things that really aren't there.

PheonixPharts··on Finding a new software developer job
I'm speaking to your point about:

> Common wisdom might make you think that I should be having companies throwing wads of cash at me

My point is just that you're misunderstanding "common wisdom". In 2016, yes a job at any FAANG would mean that you have no trouble getting a new gig. In 2024 Amazon employees do not have a great reputation, so there's no reason to think of yourself as part of a group that "common wisdom" would show preference towards. Nobody with "wads of cash" sees Amazon on a resume and thinks it's something special and worth reaching out for. FAANG in general is not as appealing as it once was, and Amazon in particular is not a strong signal at all. Someone working at OpenAI today would likely have the response you're imagining is reserved for FAANG.

Recruiters are also playing a numbers game, especially in this tech job market. My inbox has ample recruiter message for roles that are not even relevant to my skills, I wouldn't mistake that for signal.

PheonixPharts··on Finding a new software developer job
FAANG reputations are no longer evenly distributed.

Even though most people's opinions of the company have changed, Google engineers are still widely considered to be high quality. Apple engineers also have a great reputation but it's very rare to see them on the market. Facebook/meta engineers tend to be above average but have a reputation for being aggressive corporate ladder climbers.

But if I hazard a guess, if you're struggling to find work, you're at Amazon. Amazon's aggressive hire->pip pipeline has, at least in my experience, destroyed their reputation. I've worked with plenty of ex-Amazon engineers in the past few years and they were all notably below average in skill. This is not to say that you are (if you even work for Amazon), but Amazon is no longer the resume bump it once was, if anything quite the opposite.

Talking about "FAANG" is not really relevant anymore since there is a pretty big range of expectation for engineers coming out of each of those letters in the acronym. Additionally none of those companies command quite the attention they did a few years ago, as they've all shifted from demanding engineering skill to demanding the ability to survive in a large mega-corp.

PheonixPharts··on Should I use a pie chart?
> You're ignoring the reason data visualizations exist in the first place.

My opinion on pie-charts and poor visualization comes from years of studying data visualization, not ignoring it. Would you argue that Tufte doesn't value data visualization since he holds the same views?

PheonixPharts··on Should I use a pie chart?
> It's not that bad when the purpose of the pie is not to figure out which group is the biggest

But it's not better than a sorted table:

- A (90%)

- B (5.1%)

- C (4.9%)

Which group there is biggest? Which is the smallest?

> why would you not order them already inside the chart?

What's the implicit ordering of a pie chart? It's not hard to define one that makes sense (start at midnight) but it's not implicit in the chart, it needs to be explained to the viewer. Since similar sized portions of the chart are hard to distinguish, there's no way to be visually certain that you are correctly understanding the ordering.

PheonixPharts··on Should I use a pie chart?
> Charts

To be clear, I'm specifically talking about pie-charts and not all charts. There are plenty of cases where a chart conveys information appropriate to the task.

The relationships are visually represented in a table, the minimum table being just a vertical list. A table can be more, but it isn't required. Additionally the relationships in a table are more clearly represented in most all cases.

Information density also doesn't imply that you most be conveying a lot of information, it relates to the efficiency of your use of visual space.

There are no instances (aside for the other example of a clock plot) where the relationships in your data (since you don't want to use information) are not more clearly conveyed with a list then a pie chart.

> and that's why visual designers exist.

I've spent quite a bit of time working with and talking to visual designers and have yet to meet one who think pie-charts are an effective tool for visual communication.

PheonixPharts··on Should I use a pie chart?
I'll concede that this is a reasonable case.
← PreviousPage 4 of 15Next →