HNHacker News
TopNewBestAskShowJobs

chaxor

1,725 karma · joined August 29, 2020

submissionscomments
chaxor··on Recursively summarizing enables long-term dialogue memory in LLMs
For someone correctly claiming we don't know how the brain works, you seem to have a remarkable trust in the brain-gut axis pop sci articles. It's always good to remember that much of the 'established' data in neuroscience (such as the A beta hypothesis) are very well within the possibility of being based on a decent amount of fraud, as has been shown recently.

Not to say all brain-gut axis data is irrelevant, but, as with almost all of biology, the effect sizes of observables anyone typically cares about are pretty small.

However, the visual system in rodents is probably the best data source we have that maps to any formal theory right now, along with temporal difference learning and RL in neuroscience.

chaxor··on The worst programmer I know
AI is actually now shaping up to replace these jobs much more simply than blue collar work.

So, yes absolutely, administrative work now can finally be replaced, and we can free up all the tormented souls in these managerial positions to do something more meaningful with their lives.

chaxor··on Shining a light on the digital dark age
This is pretty much the best marketing for IPFS. The availability and number of backups of any price of data is directly correlated with the number of people that use it.

For example, if LLM NN model weights are distributed with IPFS instead of corporate infrastructure (basically zero redundancy) the popular models would be very available, and have essentially near zero chance of being lost.

To state that again, the llama models likely have tens of thousands of downloads, which would mean tens of thousands of servers and backups of the data, versus what we have now, which is essentially just one.

We need IPFS for data distribution. Tightly knit integration with git repos is an obvious match as well.

chaxor··on The problem of creeping commercial surveillance (2022)
Clearly can't work.

That's totally why the ozone layer is entirely gone now.

Oh wait.

chaxor··on The problem of creeping commercial surveillance (2022)
Doesn't the state just use commercial sources for their data anyway? Corporations all have far more resources and power, so it's easier to just ask for it.
chaxor··on Teaching with AI
This is also the case if taught by any educator who happens to trust the source they looked up as well. The internet, text books, and even scientific articles can all be factually incorrect.

GNNs (for which LLMs are a subclass of) have a potential to be optimized in such a way that all the knowledge contained within them remains as parsimonious as possible. This is not the case for a human reading some internet article for which they have not gained extensive context within the field.

There are plenty of people that strongly believe in strange ideas that were taught to them by some 4th grade teacher that was never corrected over their life.

While you're statements are correct in this miniscule snapshot of time, it's exceedingly short-sighted to assert that language modeling is to be avoided due to some issues that exists this month, and disregard the clear future of improvements that will come very soon.

chaxor··on Teaching with AI
You should really think about making statements like "AI will probably never do X well". Many formal linguists made very strong statements about the impossibility of (__insert feature here__, such as pragmatic implicature) to be learned by AI, which they are now being shown to be wrong.

For instance, Miles Cranmers work on using GNNs for symbolic regression is a start towards useful new discoveries in physics. Transformers are just GNNs with a specific message passing function and position embeddings. It's not hard to see that either by a different architecture, augmentation, or potentially even just more of the same, we can get to new discoveries in physics with AI. The GNN symbolic regression work is evidence that it's already happened.

As for grounding knowledge in the LLMs we have exactly just this moment (a rather short-sighted view) there is plenty of interest and work in the area, for which I expect will be addressed in a multitude of ways. It's ability with grounded physics knowledge is not perfect, but it's very good w.r.t. the common knowledge of a human off the street. External sources alone make it much better, and that's just the exceedingly short-sighted analysis of what we have today.

chaxor··on UTM – Virtual Machines for iOS and macOS
Why? I haven't ever used the iOS store, but having a git repo should befar better as an option so you can verify that it is made properly, and not infected with a little extra tracking and such whenever it's added to the iOS store.
chaxor··on Ask HN: Why did Python win?
Far faaaar better with respect to this than R though at least.
chaxor··on MapReduce, TensorFlow, Vertex: Google's bet to avoid repeating history in AI
So if I read this right, if you're not a big company (perhaps just a standard dev with maybe a tiny cluster of computers or just one beefy one), you just make a Docker container with pyspark and put your scripts in there, and everyone can reproduce your work easily on any type of machine or cluster? It seems like a reasonable approach, though it would be nice to not need the OS dependencies/docker for spark.
chaxor··on Keystroke timing obfuscation added to ssh(1)
The US hasn't really improved infrastructure since the 90s though. So yeah, it's slow, but it's also still that slow for many people.
chaxor··on Keystroke timing obfuscation added to ssh(1)
It's really hard for me to imagine that an app that markets "AI for your terminal" is going to be "more secure and private" than some standard Unix tool.

Perhaps some very specific example of a security feature (such as protecting against timing attacks) could be protected against in a new tool, and not in the older more standard one. But it seems far more likely that many other security features would get forgotten in the newer tool, and by adding "AI" so many more attack vectors would be added.

It's honestly hard to even believe in the privacy claims of warp. Almost all NLP tools in today's age seem to fall towards cloud solutions, which almost immediately makes that likelihood of privacy close to nil.

chaxor··on Beating GPT-4 on HumanEval with a fine-tuned CodeLlama-34B
It's important to frame this correctly. The article is a bit misguided (it doesn't matter which university publishes an article) because there are so many ways in which a model can be altered, even excluding retraining weights. Also, even if the performance has dropped practically due to removing some resources for more shortcuts to be taken (for example changing beam search and typical sampling parameters), making implications about the outlook for the future is not really appropriate, since retraining weights, changing architecture, etc can improve capabilities immensely.

It's important not to suggest that GPT systems in general are on the way outside.ply due to some small alterations in parameters that make a system slightly less performant (which seems to be a popular perspective).

chaxor··on Paper cups are just as toxic as plastic cups
Steel can often contain small but considerable amounts of heavy metals like lead.
chaxor··on Paper cups are just as toxic as plastic cups
Polyethylene doesn't seem that new or untested, and it's unlikely for it to contain rings, so it doesn't seem so bad. Perhaps the catalyst used that can't be pulled out after. Last time I checked I think it was a homogeneous chromium based catalyst that is used for PE, but it ends up being a very small amount left over.
chaxor··on Paper cups are just as toxic as plastic cups
Don't forget, often contains lead or other heavy metals.
chaxor··on Naked Mole Rat’s Longevity Gene Gives Mice a Longer Life
Typically Nature or high impact journals are specifically more likely to publish irreproducible studies.

It's very similar to the well established fact that students with good grades are the most likely to cheat.

If you want to look for solid research, don't look to Nature, look to the so called "mid tier" journals. Also, similar observations are true for research groups. Research from ivy-league universities follows similar principles as above.

chaxor··on How to speed up the Rust compiler
I am a bit surprised about this as well, since my perspective on a Rust comparison to C++, one of the more noticable issues is the extra verbosity rather than compile time.

You're right about switching to interpreted languages for this reason as well. Python is so incredibly simple to write, and doesn't have any compile time at all. I think a better trade off is a language that has a long compile time, is very fast and secure, but has simpler syntax.

Effectively a rust based language with simpler syntax.

I recognize some of these exist, but they need to be popular enough to use as well.

chaxor··on Code Llama, a state-of-the-art large language model for coding
This is some really good advice here. It's always a good idea to throw out all candidates that can't immediately recall what the first theoretical result of the rest mass of a Higgs boson was in the first paper describing was. Basic knowledge like this just correlates so well with ability to make proper decisions in API architecture.
chaxor··on Maccarone: AI-managed code blocks in Python
I'm fairly certain we ship far more bugs now than we ever did.

Before we had the ability to just add a patch and let the user download it, the end result needed to be very solid, because once that disk was purchased and taken home, it was static.

Now less attention is paid to these things, because it's just assumed to be tomorrow's problem.

chaxor··on FreeBSD replaces bubblesort with mergesort on SYSINTs
"This is good for LLMs"
chaxor··on Vector databases: analyzing the trade-offs
This little thread helps solidify my opinions towards using only the few in the small section they point out, being Qdrant, chroma, or weaviate. The fact that someone who works at pinecone didn't take the time to read an article in their field, decided to comment anyway, and then stated they didn't know what serverless or embedded may mean with regard to databases pretty much sums up the state of many vector databases, especially pubecone imo.

I don't even really love Qdrant or chroma (haven't tried weaviate, but it's at least in the right region according to the article), but at least they are embedded.

I pretty much refuse to use any DB that requires using API keys, putting data off premises, and even if it requires setting up ACL. I don't even use postgres much for the complex ACL and having to set up ports reason.

SQLite and DuckDB are truly incredible, can store gigantic databases (>2TB is still perfectly quite performant) and you can just hand any collaborator the entire DB on a disk, without having to worry about complex password junk.

chaxor··on Vector databases: analyzing the trade-offs
It was my understanding that txtai is not a vector database, but rather it uses databases. It pulls together a large set of tools that most people use together in a very nice way, such that the API is more consistent for researchers to show what they are doing.
chaxor··on Run a macOS VM on Apple Silicon from a double-click with Vimy
You mean qemu.

Utm is qemu, but they just removed the open source part. So it's just definitely worse.

chaxor··on Run a macOS VM on Apple Silicon from a double-click with Vimy
Utm is just qemu, and qemu is actually open source. Utm hides an enormous amount of what they do (just saying not open source again in a different way), and that's definitely not what you want in a base system.

So it would likely be much better to simply install qemu instead.

chaxor··on IP address blocking banned after anti-piracy court order hit Cloudflare
It's mostly navy run right now is my understanding, and used by journalists, so I don't see how it's that much of a downgrade.
chaxor··on Show HN: Aviation navigation log on $20 receipt printer
You still have to fill it up by touching and breathing in the fumes while pumping it into the tank, so not exactly, but you're right that it does affect others as well.
chaxor··on IP address blocking banned after anti-piracy court order hit Cloudflare
I always assumed businesses such as Cloudfare or Google would have a good reason to be the main IPFS or tor node hosts, since that can help improve the backbone infrastructure for this side of of the internet, and can also help catch bad actors if there are any there (easily turn in all logs to three letter agencies if they knock, or they can simply offer it up for anon tips).
chaxor··on Why do old books smell so good?
I love the smell of doing a good wipe on my hardware by sticking all of it in the microwave.

The smell of burnt silicon in the morning is what keeps me waking up everyday.

chaxor··on Show HN: Aviation navigation log on $20 receipt printer
General Aviation still uses lead in their gas, despite it being shown several times in enormous nationwide studies consistently to not be necessary, and often simply requires more cost in maintenance (lead builds up on spark plugs causing a lot of issues).

So environmental impact isn't high on their list already, and this BPA issue is very likely going to be completely ignored by anyone in that field.

← PreviousPage 7 of 31Next →