Hacker Newsnew | past | comments | ask | show | jobs | submit | hypendev's commentslogin

I'd rank them like this - this comes from applying LLM's since GPT-3 autocomplete API, teaching people how to use them, giving talks to developers, answering their questions, training my own models and seeing how people interact with them:

1. Building your own model - this will cost you a bit, but is without a doubt the best way to earn concrete AI skills. From picking the data to training, seeing how it evolves, improves and scales. You will understand the "hyperplanes" of a model's inner world better than anything else can teach you. You will learn how to pick data, how it impacts model quality, way of thinking, scores, personality. If training your own is too much, start with a completed pretrain and toy with it.

2. Building a harness around a completion model - I'd put this as second best to understand the inner worlds of LLM's. It helps you build intuition around how prompting really works, how they form the answers, how to build around that. Then you can do an SFT for your own harness.

3. Building a harness and an orchestrator around finished models - this takes the bronze. It teaches you how harnesses, agents and swarms actually work, how they shape the model's thoughts and actions, how models interact with them, how they can be applied and how they can work without humans.

.


How realistic is it to build your own model from scratch costwise (in terms of training). Any suggestions as to what is realistic with e.g. a decent MacBook Pro?


Well, depends how much time you're willing to invest.

A small model, i.e. a few hundred million params might take you a few days of training, while bigger ones (0.5-1B) would take you months probably.

But you can go for it until you reach the scale where the time tradeoff isn't worth it and use kaggle, colab or such free GPU resources to continue trianing it.


Personally I think training an LLM from scratch locally is a bit pointless. It's too difficult a problem. Your results will vary from a model that can't do anything to a model that can barely do anything – and that's if you invest a significant amount of time in it.

I actually think just learning about the fundamentals of neural networks and building your own models to solve somewhat simple image recognition or decision problems is more useful and more informative.

That said I basically stopped learning about any AI stuff a year and a half ago. Realistically humans are only going to have an edge over AIs in AI development for a few more months now.


Ironically, did the same thing and ended up with the same answer.

From wearing a CGM, whoop, using an impedance analyzing scale (I know, I know), blood tests - all the data shows pretty much the same correlation. I mostly did it for fun, but also because I wanted to "figure out how different things impact my body".

While yes, there were some interesting discoveries like: - Some things raise my blood glucose more than I'd expect, pasta for example - those were the things I usually "knew" and foods I'd avoid eating.

- Sleep quality only had correlation with amount of activity/time I went to sleep, so my food and caffeine choice didn't impact it much.

The results were pretty much: Eating bad and moving less is bad for you. Eating good and moving more is good. What your intuition tells you is healthy and good for you, usually is. What it says isn't, usually isn't.

Could have saved me some time and money, but at least I had fun.


I'd disagree with the statements and people making fun of it.

While these people are obviously professionals in their field, anyone who touched pharma/bio IT with a ten foot pole understands how much there is to improve there. And that's just the beginning - imagine how much there is to improve in understanding of the body overall?

It's not about "AI will discover a cure for X in 10 years".

But it's about having the applied intellect of a an engineer, chemist, biologist, and all mixtures of that, available on demand, in a second, at scale, enabling you to do everything around it faster, better and cheaper.

Every step of drug discovery, synthesis and testing can be improved.

Over 90% drugs tested in humans fail. And maybe 0.1% of overall discovered ones. While yes, testing on humans is the bottleneck, we can fix problems around the bottleneck easily.

If you had a data processor and 90% of data that goes into it is corrupt, the processor isn't the bottleneck yet - the data quality is.

If we can create better biological models to test these drugs, better interaction models, improve pharmacokinetics, improve yield, we can go from 1000 discovered, 1 works, to 1000 discovered, 100 works. We can predict some candidates will fail due to better models and not test these. We can improve how they are delivered to the body and make them more effective. We can catch potential side-effects early on and avoid going to trials at all.

This is what AI can do even nowadays, let alone 5 or 10 years from now.

And that is a huge improvement. It revamps the industry.


the data quality is the bottle neck, so why do you think improving the processor will suddenly unbottleneck it?

tbe improvement you are suggesting isnt whats being sold - that ai will actually discover said drug


It isn't - what Dario is saying, exactly, is that these diseases will be a thing of the past soon and humanity, with the help of AI, can solve them.

Things that are being sold overall are exactly those:

- Helping identify better candidate drugs

- Building better models for finding and eliminating them

- Startups using AI to build lab-grown human tissue

- Using AI to manage thousands of microarrays and test out different permutations of same candidate in parallel

All these things will boost pharma massively.


> the largest theft of intellectual property in human history. Every digitised creative work, whether in the public domain or not was taken, often outright pirated, with the explicit intention of being able to trivially produce compelling forgeries

Is it theft?

Does knowledge belong to a person, or to the civilization? After all, their knowledge could not have come from nothing, as they too stand on the shoulders of giants - all of us do.

If we took the different route - opening knowledge up instead of focusing on gatekeeping - we'd probably allow people to achieve more.

Multiple times I've hit the gate when it was not necessary. Just yesterday: - Want to read a research paper? Pay up. - Want to access tax funded satellite images? Pay up. - Want to dig through some old parts of the public archives? Pay up. - Want to read an old publication that might contain knowledge, written for the public, which somehow ended up in a collection that requires you to pay up just so you can read 1 paragraph from a 188 paperbck or verify a certain catalogue number exists? Well well, open up your wallet.

In which case are we talking about theft?

Someone's NYT bestselling novel? Those are mostly so cookie cutter it ain't hard to copy them without an AI.

Music? How many times have musicians riffed and ripped off of one another? And now someone can just more easily create a similar song/tune.

Art? You cannot copy a physical art piece. You can reproduce it, but you cannot copy it - it won't have the same value as the original. Even in cases such as Warhol's mass produced art, the original prints have much more value than a random print on a dollar store t-shirt.

And after all, even Google and Open AI are sharing their models for free. You can use Gemini and ChatGPT without paying a dime to any of them. And they are also so kind to share the knowledge on how to build these magic machines AND even give you their machines for free so you can train one of your own.

So I wouldn't call all of this "the largest theft", more like "the largest liberation".


> Is it theft?

Well we certainly didn't get paid for it, and now our livelihoods are at stake. So yes.


Who's?

If you're a writer whose work people enjoy to read, they will still read it.

If you're a painter whose paintings people enjoy to see, they will still look at them.

If you're a musician people enjoy listening to, they will still listen to you.

Nobody's money is being taken here. It's not like people are prompting "Give me xyzsparetimexyz's book cover to cover".

Only think that's taken from someone is the idea that they have the right to own a piece of information.


> It's not like people are prompting "Give me xyzsparetimexyz's book cover to cover".

They are literally doing that with slop game clones trying to beat them to market.


Even before AI you had a swarm of slop clone games. Half of steam and itch are "indie" games with are just absurd slop or a game clone.

And still, these games won't be played or take up the same space that the originals have in players minds. I'll enjoy playing GTA 6 - I won't enjoy playing Vibe Car Theft 6 which emulates it. Especially in indie gaming, where people relate to the creator and choose to support them.


you're basically saying "it's not theft, because now it's mine"

also,

> Someone's NYT bestselling novel? Those are mostly so cookie cutter it ain't hard to copy them without an AI.

lol then why weren't you a bestselling NYT novelist, had something better to do?


> you're basically saying "it's not theft, because now it's mine"

No, that's not at all what they're saying. Taking something is only theft if it's rightfully owned in the first place. He's saying it wasn't rightfully owned.


Yes, I find other stuff much more entertaining than writing cookie cutter novels for bored housewives and pseudo-intellectual pop-sci books for dinner conversations.

But also, none of the ideas in these books are novel. Is it intellectual theft to be trained on a book that is based on public research where author just distills the information into an easier to consume format?


Oh god I dealt with this recently while training my own 1B model.

Runpod - mostly okay, except sometimes instances get dropped on smaller ones. Prices float around 3.20 mostly for H100s. Cheaper ones exist but are PCIe. Best DX out of the bunch honestly (well, AX, as my agent was the one dealing with it). Didn't end up doing the 1B run on it, but did 100M+ runs and some minor experimental ones. Great for agentic exploration, "RSI", minor training loops and research.

Vast.ai - a bit pricier for the same H100s and it didn't have any clusters. So I could at best split my workload across 2-3 different H100s and spend time syncing. Not worth it.

Akash - since I'm trying to stay in a reasonable budget, I decided to try Akash, which was the worst experience out of all. Cards would be "available" but would never respond to your request. After trying to contact support via Discord, I got dozens of fake support accounts adding me to scam discords and asking for my wallet key. Support answered after a while, but without a real answer.

Mithril.ai - Ended up using this one, while it seems quite startupy, the actual offer is nice, it was easy to get going and they offer quite cheap H100s, offering singles and clusters.

Tbh while Skypilot is great, the marketplace fragmentation is leaving room open for a openrouter-like compute provider.


This is why ClusterMax (from Semianalysis) was created. Not all hosts are the same, even with the same GPUs.

https://newsletter.semianalysis.com/p/clustermax-30-the-indu...


Oh nice, did not know about this.

Looking at the report, I didn't even know one could use CoreWeave without an enterprise account, seems it could be cheaper per hour than the same cards elsewhere. Might have saved me a few hundred bucks, thanks!


I was confused a bit where this was. Then I had to Google Gemini, which lead me to gemini.google.com (did not know it existed, and I use gemini in prod via google cloud).

First thing there that got me was that all of my Google searches. Why? I googled "snoop dog real name". Why is this considered a gemini chat now? If I actually used it, my chats would be buried under dumb quick google searches I did. And it seems its only my phone searches, not alL?

I assume because some PM will say "we have 1243153344% more Gemini searches than last quarter", then gonna get paid 3 million USD bonus for something that ruined the UX.

Then, I looked at the top and didnt find the model dropdown. It is next to the input, okay. I click it. Argon? Locked. Even tho I pay for the API anyways and can use it via the API.

But lets continue.

My app stopped working 2 days ago for no reason - apparently, I needed to move to a "prepaid payment mode", which wasn't a problem for the last 2 years.

- I open the AI studio, the banner to update shows, then page reloads and it disappeared.

- I was like okay, lets see - checked billing, tried to top up. First time my bank's 2FA notification didnt come, so I rejected and tried to top up. Then it would reuse the same windows as before with rejections. Okay, I refresh, topup, it works. I try to open usage to see my usage, but I can't see the spend there in any currency besides tokens.

- Now let me try it out. I type in the prompt, select the model, hit enter. Oh wait, I have to link the API key. I click the button, a dialog opens, then closes itself without any information. The prompt? There is no prompt anymore, you have to type it out yourself.

While I believe in Google's AI offering - I use their models and for my niche they are the best for good quality at great speeds.

But holy shit, every time I use one of their products, I just get annoyed. While Logan and AI studio team have really done some big improvements, the overall Google offering is confusing as hell.

Why does a customer have to hear "Oh yeah, don't use Google Cloud now, use Vertex via Gemini Enterprise Agent Platform"? Because some PM chased a bonus?

The internal politics are definitely ruining any edge they might have.


> First thing there that got me was that all of my Google searches. Why? I googled "snoop dog real name". Why is this considered a gemini chat now?

It isn't - it's probably an _assistant_ query? In fact the Gemini and Search orgs are cartoonishly at loggerheads; both developing more or less the same product (http://gemini.google.com vs http://google.ai).


No, it is a search from my launcher on S24. Literally google searched from the launcher as I always do. I specifically remembered that, because I was sitting on the couch with my GF watching a movie and quickly googled his real name.

Now that I look at it, weirdly, both show different results.

Gemini shows only what I searched on my phone, meanwhile Google.ai shows ALL of my searches as chats.

Guess both PM's are getting a big bonus this year, and the stock price will get a nice bump from "Look at how much people use our AI".


Weird, maybe it's specific to your launcher? My searches from the -1 page of the launcher don't appear in Gemini, and the ones from the bottom corner swipe to search or OK Google don't appear in AI mode.


Congratulations to the winners!

From my understanding, this discovery is on the level of what transistors were for electronics - unlocking a completely new way to understand and design the electronic circuits, which later on enabled the rise of computing.

Since compared to biology our computing is still highly primitive, it seems like next step in compute power will be biological and quantum computing.


It could be in the future but its not yet. Optogenetics has just merely been one more albeit powerful tool in the neuroscience toolkit and have generally not been impressed with how useful (or not) it has been..


I think its hard to measure the impact directly, but I'd say it has an enormous indirect effect on neuroscience that is bigger than anything yet I think.

You have to imagine, before opto, EVERYTHING in neuroscience was either genetic or observational, nothing really good with manipulating circuit function directly except for the only real intervention we had - ablation of an entire region without any specificity, especially with regards to passing axons.

Now, I literally expect any neuro paper that even touches behavior or circuits to have some sort of validation with opto, or it's harder for me to believe.

Maybe there's a bit of an over reliance on it in my above comment, but if you have a better causal tool to suggest, be my guest.


Good choice of username.


Exactly this.

Streaming services have become the worst of all worlds.

As an EU citizen, even tho I pay for Netflix and Prime, I cannot access most of the content there as even tho it exists on the service because it's not available in my country.

Amazon prime has 2 good shows and just a bunch of D-tier movies or some 35+ year old ones. It's mostly worthless.

Netflix region locks a whole lot of content or removes it after some time. I can't watch some great movies and TV shows because "oh we don't have rights for this region".

Meanwhile, piracy provides a user experience that not just lets me access all of their content - one I'm already paying for - anywhere, but also lets me access movies and shows that would otherwise not be found anywhere else, including regional movies and short movies.


Content rights are micro managed too, unlike ever before so you get things like a streaming service buying a title in one language but not the one you want. This doesn't stop them from returning it in your search results though


I watched Star Trek TNG on Netflix like four times.

I have never subscribed to Paramount+ and people who liked TNG are saying things that make me think that was the right choice. I just don't watch TNG anymore. That's fine, it's not like content is scarce. Good content is somewhat scarce, but there's decades of known good history to dig through.


Netflix had an advantage with own productions which is dying now. Shows are cancelled left and right so people don't want to invest time. Also looks like full season dumps is affecting the long term profitability of productions. People watch a season in a couple of days and forget about it, if a new season is planned it takes 2 years to be released by which time viewers forgot the show existed.


Prime's local content issue can be fixed with a VPN.


Doesn't seem to work for me, always shows the same


Ah yes, always use a vpn when torrenting.


Having worked in/with multiple agencies over the years - working with humans who don't care/are pressured to just ship was exactly what he describes.

For example, I collaborated together with an agency that was working on a popular regional banking app. I was on another project with them and at one point decided to check out what they were doing.

It was a clusterfuck, to say the least. Four developers, four different architectures in the same codebase. One doesn't really understand the others, so he just sticks to his part.

The seams between the ownership parts are full of hacks. Concurrency issues are dealt with via hacks, just make things global and poll to check them. Navigation issues were solved by basically just reseting the app back to homescreen. Testing wasn't really a thing, and the tests that were there were mostly broken ages ago, and nobody cared enough to fix them.

Product didn't care much, designers would prototype a feature, then devs would implement it and wait until the backend is done, then fix things because they never really aligned on the specs. Sometimes, there would be no design review or QA and they would just ship, while sometimes the product person would jump in and test something because it's "an important feature".

Then I got asked to collab with them on another project. Huge customer, huge project, existing codebase. The codebase was so bad I just wrote it from scratch. It was Java written like it's C++. Backend was python code, not updated in years, quadratic scaling everywhere. At one point, the inital GET request was taking over 3 minutes. I took a look and ofc, they didn't even use indexes in the DB.

For these environments, AI coding will just scale up the shit they have, but also allow the small amount of people who might care to deal with this mess and fix it up much much faster than they could before.


> For these environments, AI coding will just scale up the shit they have, but also allow the small amount of people who might care to deal with this mess and fix it up much much faster than they could before.

I think in these contexts AI has an overall positive effect, because it's a bit like having a competent developer that can solve technical debt, albeit in a narrow form.

So I think that software will be better in a certain way due to AI, because even a terrible developer can ask AI to identify the bottlenecks and solve them.

The downside of blindly using AI in software development is that the software becomes poorly defined. AI-generated logic can be arbitrary, and as the codebase grows, more rules and assumptions become unclear.


> So I think that software will be better in a certain way due to AI, because even a terrible developer can ask AI to identify the bottlenecks and solve them.

The root cause of the issue is not the bottlenecks or the ability to identify them. It’s about caring about doing a good job. Because once you start caring, you will need to exercise judgment. Better or worse only matters when you have goals.

This is one reason we have so much slop with AI. The cost of the journey has lessened, but you still need to have a destination and be willing to appreciate the journey to get there. Without, it’s just endless drifting.


Depending on the plan and amount of work. Bridgeflare stops your containers after idling for 10 minutes then boots them up again once needed (to avoid burning money), so if you're already on the 5$ plan and doing normal workloads, it should fit in the plan.

If you're running them live 24/7, then costs can be in 40-70 dollar range, depending on the workload done/type of instances.

All of my usage last few months has been in the 5$ plan, as they go idle as soon as I stop testing and come back up when I need to test.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: