HNHacker News
TopNewBestAskShowJobs

icyfox

1,616 karma · joined April 6, 2021

pierce [at] freeman.vc despite the email i mostly code
submissionscomments
icyfox··on How to ramp up a factory consuming a lot of energy?
A small - and somewhat related - anecdote.

New Zealand had to build an aluminum plant in the 1960s and figured out it would be more efficient to bore miles underground, build turbines, install a huge power station, and wire it tens of miles to the smelting plant. It relies on a vertical drop between Lake Manapōuri and the open ocean to create the gravitational potential to turn the turbines. When the smelting station doesn't use the full capacity the power station has to immediately reduce output, otherwise it can overload the transmission lines to the rest of the grid.

Manapouri Power Station is the name if anyone is interested in reading more. It has an interesting history. https://en.wikipedia.org/wiki/Manapouri_Power_Station

icyfox··on Show HN: State Trooper – Tiny, no frills state machine for Go
Midjourney actually does pretty well with this one (needs a bit of cleanup but shockingly useful as a first draft): https://imgur.com/a/GKzx3vX
icyfox··on Show HN: GPT-JSON – Structured and typehinted GPT responses in Python
You might not! Depends on what you're looking for. I've been finding this library most helpful in places where I have a lot of GPT calls in pipelines, so having typehinted schema return values / some built in error correction / variable injection / establishing standards for the IO of prompt schema is the most useful. So IMO I see its main use as a good standard set of operations that work pretty well out of the box and that allow you to hack around them with decent flexibility.
icyfox··on Show HN: GPT-JSON – Structured and typehinted GPT responses in Python
Good suggestion mmaia - I'm opening a new issue to keep track of the different output schemas that are being suggested. One thing that originally worried me a bit with yaml was its relative reliance on space-based formatting to drive meaning. GPT generally tokenizes newlines/spaces but a lot of preprepared datasets strip these out, so I preferred the explicitness of a json that's idempotent to spaces. Have you tried using this approach for non-tabular or nested data like lists or dictionaries?
icyfox··on Show HN: GPT-JSON – Structured and typehinted GPT responses in Python
Pretty good pitch from ChatGPT if you ask me! I haven't used cue outside of test projects before, so a few questions:

1. Is cue's validation a material improvement from something like pydantic or zod, which defines schema as code versus in .cue files? I see their docs argue that this can allow for client-side validation and lighter weight schema files which doesn't seem to totally address the library side of things.

2. Have you used the scripting layer before and do you find it useful in practice? I'm struggling a bit to see how I or GPT would use this in my day-to-day.

icyfox··on Show HN: GPT-JSON – Structured and typehinted GPT responses in Python
Fingers crossed OpenAI / Anthropic / etc do this! Would make working with these APIs for prediction projects that much easier.

Technically speaking it's pretty to force the model into an valid JSON-schema if you have access to the inference autoregressive loop and the logit activations. You can either force known areas of the template to a prefixed template, or fill in the basic JSON wrapper and force the model to choose arbitrary keys and values.

Imagine it might come down to how much of their usage is on generating text vs generating structured prediction payloads.

icyfox··on Show HN: GPT-JSON – Structured and typehinted GPT responses in Python
Thanks for the thoughts! I've deployed a few meta models that act like you're describing for second stage predictions, but for fuzzy task definitions have actually seen similar luck with having GPT explicitly explain its rational and then force it to choose a true/false rating. My payloads often end up looking like:

  class Payload:
    reasoning: str = Field(description="Why this value might be true or false),
    answer: bool
Since it's autoregressive I imagine the schema helps to define the universe of what it's supposed to do, then the decoder attention when it's filling the `answer` can look back on the reasoning and weigh the sentiment internally. I imagine the accuracy specifics depend a lot on the end deployment here.
icyfox··on Google “We have no moat, and neither does OpenAI”
One issue with the current generation of open source models is most have been based on some llama core architecture, and that's not licensed for commercial use. Once you get to the point of spinning up a full and easy API, and selling API credentials, you're entering into the commercial clause. Once we have a llama alternative (or a more permissively licensed separate architecture) I guarantee hosting providers like Render or Model are going to come in with an API offering. Just waiting on those core models to improve licensing, would be my guess.
icyfox··on SparseGPT: Language Models Can Be Accurately Pruned in One-Shot
Another random thought: Most of these general purpose pruning approaches rely on randomly calculating the X vector for which they want to measure output loss of the layer. In theory it's possible to feed actual datasets into these models as well, which could be another way to get a sparse model that's more acutely optimized towards one task. The original model produces the X activations on each layer, and these are used as the optimization criteria for the pruned version.

It might be able to provide performance similar to fine-tuning but without the weight skew that you'll necessarily see in parameter values.

icyfox··on SparseGPT: Language Models Can Be Accurately Pruned in One-Shot
Neat paper. Planning on reading more in-depth over the weekend, but more fundamental than just applications to GPT their insights are:

- Existing pruners were written for models that are order-of-magnitudes smaller than any in the modern GPT family. They grow in linear time with the amount of input parameters so they're unequipped to work on current architectures. The best existing pruner performs takes 4.3h for a 1.3B model

- The core issue to scale is time to calculate the Hessian during prune analysis (effectively a matrix of second-order derivatives, famously computationally intense to calculate)

- They follow the existing literature and use a local approach to each layer. By doing this (and doing it well), it can preserve the input/output contract for surrounding layers, which makes the whole thing paralellizable across machines

- Their solution approximates reconstruction loss by approximating a quadratic loss and then running a OBS update (with a few other optimizations on ordering and iteration on the side)

I'm particularly excited for these smaller models, mostly for inference efficiency gains in realtime applications. The general con of weight pruning is they still require incredibly large training clusters / investment in training resources upfront to get the original parameter weight. But if the lottery ticket hypothesis holds true, this might be the best way we have at the moment to get models with same performance and lower longterm operational costs.

icyfox··on OpenAI has applied for “GPT” trademark with USPTO
They're going to have to clear a pretty big hurdle in establishing a trademark once it's been out in commerce for so long.

But let's assume that they do. I actually think this will prove to be a good move for the ML ecosystem over the long term. There are so many products branding themselves as "[Vertical]GPT" that many non-technical users are under the illusion that the solution is endorsed by the same people that make ChatGPT. People in the tech circle knows that OpenAI isn't pushing out "DogBarkGPT" but with the shared suffix that's not immediately obvious. Their suggested "powered by GPT" is a much more honest phrasing.

While being too aggressive about brand guidelines and usage patterns can unintentionally harm innovation and competitors, I don't think encouraging others to stay away from capturing the phrase credibility is a bad idea.

icyfox··on US National Debt Tops $31T
In the government (and punditry / public discourse) it sure seems to be. Academic and financial circles only matter as much as they can sway the former to make policy.
icyfox··on US National Debt Tops $31T
I'd like to see some compelling evidence that supports the seemingly prevailing economic theory that the US can issue an unlimited amount of debt just because it's denominated in USD and interest rates are historically low. From another NY Times opinion piece (in [1]).

> And the new paradigm suggests both that public debt isn’t a major problem and that government borrowing for the right purposes is actually the responsible thing to do.

> Here’s where it becomes crucial to realize that the world has changed: Interest rates are much lower than they were in the past, and all indications are that they’ll stay low for years to come.

The phrase "for the right purposes" sounds like a scape goat for government spending. What is a right purpose? Morally speaking - almost anything. Economically speaking, I imagine only things that are likely to increase net GDP over the long haul.

It seems clear to me that you can draw a pretty clear line between quantitate easing & government bailouts of companies during COVID to a sizable quantity of current inflation. Asset prices rose in the market, which bubbled down to housing, rent, and consumer products. Supply shocks for oil and trade commodities were part of the equation but only a part. From everything I've read on federal bank history, the more money that's in supply means the higher inflation - almost by definition. You can only issue so much debt out of nothing before dollars are fighting over the same core commodities.

[1] https://www.nytimes.com/2020/12/03/opinion/biden-republicans...

icyfox··on Boxes.py – Create boxes and more with a laser cutter
I first used a laser cutter ~eight years ago. Our printer's UX relied on sending designs with special colors to an attached usb driver. So you'd make something in Illustrator, open a regular print dialogue, and get out a precisely cut acrylic piece a minute later. It was pretty incredible.

This was back in college so we had a manufacturing lab that was chock full of them. They're a bit harder to find in the wide world - but highly recommend the experience if you can get your hands on one.

icyfox··on Why none of my books are available on Audible
Audible has some flagrant dark patterns in their design for consumers as well. While you _could_ buy independent audiobooks at cost, their most popular model is to subscribe to a monthly membership where you earn one credit a month. This monthly credit can buy any audio book.

The dig comes if you decide you have too many credits cached. Perhaps because you can't get through one audiobook a month. If you cancel a subscription for an account with remaining credits, you'll lose all your credits. Your choice is to either spend them all at once or keep your account active and keep paying the monthly fee while hoping to eventually pay down the credit debt enough to cancel.

The only thing that's compared in my experience is for Adobe Stock.

icyfox··on Falling for Kubernetes
OP here. To clarify, the blog doesn't use containers just because it didn't really call for it. Ended up creating a simpler CI pipeline to just push updated static files directly. For everything else that I work on, I'm 100% in on containers. Just those containers running mainly on bare metal and newly on k8.
icyfox··on Ask HN: What core research are you most excited about?
Are we materially further along on fusion than we were 10 years ago? I know we're making some strides with magnetic / vacuum sealing but I'm unclear on how much novel insight the research has surfaced, versus just iteratively improving our ability to physically prototype it.
icyfox··on Tesla driver using Autopilot kills motorcyclist, prompting another investigation
Even if we accept some statistical failure rate, we should still investigate lethal failures. Like any good RCA it may uncover some learnings that could help the implementation and/or regulations around self driving.
icyfox··on The most satisfying checkbox
No - but how many habits are you trying to form at one time?
icyfox··on The most satisfying checkbox
Depends on how frequently you're interacting with this widget. The annoyance of animations comes from the integral of all time you have to spend watching them; if they're in the critical path of a workflow that you do multiple times a day they get very annoying very quickly.

I've only glanced at this app, but the goal seems to be forming a habit of doing things once a day. If you're only interacting once every 24 hours, I'd take a 1second animation with some charm over a static checkbox.

icyfox··on Apple Unveils M2
I can't figure this one out either. The Air now has the same screen as the MBP 13", same processor, slimmer form factor. The air is also $100 cheaper than the MBP so that's not a justification. The only thing that is different on the specs page is the GPU count 10-cores versus 8. Why keep the 13?
icyfox··on Hiring as an Early Stage Startup
Are you the alumni of any other companies or engineering schools? I've had to hire for similar sized startups and find the most success with reaching out to past colleagues or friends in the same space, and either pulling them in directly or having them extend a quick pitch to the people within their networks.

Until you get to ~15-25 engineers, this has been my highest yield way to hire.

icyfox··on Launch HN: Slai (YC W22) – Build ML models quickly and deploy them as apps
Congratulations on the launch guys. Product need seems clear to me & is a painpoint that I've felt most acutely in side projects that I've worked on outside of our company's devoted CI infrastructure.

Are you planning any git or IDE integration? Most of the magic here seems to happens in the backend with easier training, scheduling, and inference. Could this be enabled locally so devs iterate in an environment that's more comfortable to them?

icyfox··on Brrr.money
Right when I feel like the internet's becoming too corporate and optimized towards SEO for the sake of profit and profit alone, I land on a gem like this. Thank you brrr lords for bearing the webhosting cost for something fun.
icyfox··on DMV approves Cruise and Waymo for commercial service in parts of Bay Area
Waymo has a similar approach in Phoenix. Passengers can't directly intervene but can call tech support and they'll dispatch roadside assistance to override the car and drive it to safety. Will be interesting to see if Cruise copies this same model.
icyfox··on Private censorship is not the best way to fight hate or defend democracy (2018)
I define a radical group to be some collection that seems to be consistently peddling false material for their own ends, with an overly provoking tilt that makes it easy to go viral. Some specifics out of the past year would be vaccine misinformation, prejudice that incites racial violence in Myanmar, organizing an occupation to the capital in the US. Shouldn't we put a stop to these communities if they sway people to start peddling this information themselves?

At least in the US, free speech is most free within public forums. But even then we already define some speech that is too dangerous if it's known to lead to poor outcomes. You can't yell fire in a crowded room, not because yelling fire is inherently a crime, but because it's going to incite people to a detrimental ends.

Plus social networks don't have a constitutional obligation to be town squares where free speech can spread unfettered. You have to draw a line somewhere. And as the people who write the algorithms that can amplify or bury content on these networks, I think it _is_ our obligation to at least set parameters on what constitutes good/healthy content interactions on these social platforms and what doesn't. The ML algorithms have to optimize over some loss function.

icyfox··on Private censorship is not the best way to fight hate or defend democracy (2018)
These things do help. Most members of these communities get radicalized by virtue of other content they’re already browsing. If you remove the communities and force individuals to leap from Facebook to some no-name forum, they’re far less likely to engage.

This phenomenon is mirrored elsewhere as well where small interventions can lead to big impacts: see suicide rates in Britain after removing carbon monoxide in their house gas. Initial friction (or means reduction in that research) can drive change.

icyfox··on Dithered images and websites (2020)
Images serve to make a website beautiful. Dithered images are not - in my mind - beautiful. There are other ways to optimize a website speed and bandwidth. Let's serve devices the image resolution that they can handle and not an XL image just because we can. Let's minimize JS cruft, minimize excessive renders. Sacrificing image quality would be near the bottom of my list.
icyfox··on Airbus reveals the next generation of CityAirbus
As much as I want to live in a world with the Jetsons, what's the business case here? Presumably it's going to have to launch vertically from the top of buildings, which means that you can go from skyscraper landing pad to skyscraper landing pad. How many taxi users are trying to travel within a financial district? I imagine much more commonly they're trying to go from high-density to low-density (like FiDi to West Village) which this operating model could only support if you build a serious amount of elevated architecture.

Not to mention having to pay for insurance premiums of a miniaturized airplane that is zipping around a heavily inhabited urban area.

icyfox··on Ask HN: Who is looking for a co-founder? (July 2021)
Makes sense.

A lot of my work involves parsing crawled pages, ie. html -> structured dom tree -> some heuristic search or other algorithm to tag or extract elements. So in this case the compute per page is relatively high but I still think the overall latency of having to sync to a remote endpoint is too heavy.

One hybrid approach would be some hosted data product that is colocated with your compute cluster. On nonsensitive datasets I'd be happy to upload my initial data upfront if it means I can get linear scalability when I need to run compute.

← PreviousPage 4 of 5Next →