Hacker News Reader

byefruiton Show HN: Wordllama – Things you can do with the token embeddings of an LLM

What are some recent sources for high quality similarity pairs?

byefruiton Ask HN: How can I monetise an open source app of mine that has become popular?

How do you handle contributions to the open core? Do you require a CLA?

I've been wondering about this for a project of mine. If one goes for an AGPL license but offers a commercial license for those that can't comply then a CLA is necessary from any contributors.

byefruiton What happened to BERT and T5?

Yes, no zero shot. Few shot is possible for some use cases with setfit: https://github.com/huggingface/setfit and the very recent Fastfit: https://github.com/IBM/fastfit ( https://arxiv.org/pdf/2404.12365 )

byefruiton Logseq – adding settings for self-hosted sync

100% agree. If users want to remove a feature that funds maintenance and development of the project then they should really just fork it.

byefruiton Gpu.cpp: A lightweight library for portable low-level GPU computation

Thank you!

byefruiton Gpu.cpp: A lightweight library for portable low-level GPU computation

This looks great. Is there an equivalent project in rust?

byefruiton Falling in love again with disposable cameras

Exactly. There's no point stopping unnecessary and wasteful practices until all other unnecessary and wasteful practises have ceased first.

byefruiton BYD Rolls Off Its 7Mth New Energy Vehicle

In relative terms (of new cars sold) they're down around 1%. That isn't good when share needs to be increasing.

Policy support for EV adoption has been woefully poor and inconsistent (you could generally say this about most things from the government at the moment...).

It's hard to make investment decisions around EV infrastructure when the government keeps changing their plans every leader.

byefruiton When 'open core' projects reject contributions for competing with the EE

This project is MIT-licensed, so that applies here too?

byefruiton When 'open core' projects reject contributions for competing with the EE

But you've always got the right to fork it.

What you're saying is you want them to continue maintaining it at their cost because the community don't want to?

byefruiton When 'open core' projects reject contributions for competing with the EE

I don't see the problem here? It's their project and they can choose what they want to maintain. If you disagree, it's perfectly fine to fork the project and implement this but then you bear the cost of maintenance.

These kinds of open source projects need to be sustainable, you can't expect the companies producing them to continue maintaining and investing in them without some sustainable way of doing it.

I don't see why project maintainers gets called out on this but AWS reselling open source projects whilst contributing very little to them isn't.

byefruiton Review: Framework's Laptop 16 is unique, laudable and flawed

This mirrors my experience with Framework tbh. I love the hardware but they've let us down at pretty much every stage when it came to customer service.

First laptop clearly had no QC and instead of offering to swap it, the expectation was we would do some pretty tricky repairs to a brand new laptop. We sent that one back but the return just sat there for weeks without being processed before a post on the forums got it moving.

The other laptop refused to boot after eighteen months (still under the two year EU warranty) but involved about twenty back and forward emails to support - half of which ignored diagnoses we had already done and provided the results for.

I love the hardware but I just don't think I can recommend people buy one at the moment, which is a real shame.

byefruiton New embedding models and API updates

Probably something similar to https://arxiv.org/abs/2310.07707

byefruiton Fireworks: Function Calling Model and API

Interesting but if you're expecting an actual OSS model, it's only usable through their API.

byefruiton GateLoop: Data-Controlled Linear Recurrence for Sequence Modeling

Lucidrains has a re-implementation of this: https://github.com/lucidrains/gateloop-transformer and was unable to beat the transformer baseline at equal numbers of parameters.

byefruiton All my favorite tracing tools

I think undo have one: https://undo.io/products/java

byefruiton One million cancel broadband as living costs rise

https://ifs.org.uk/taxlab/taxlab-data-item/combined-marginal...

49% kicks in at £55k at combined NI and Income.

(This all ignores the 14% that's hidden on the employer side)

byefruiton OpenLLaMA: An Open Reproduction of LLaMA

They ran lm-evaluation-harness on both this model and the original llama weights, which is the correct way to do it.

Many people have been struggling to reproduce the benchmark numbers included in the original llama paper.

byefruiton Framework’s computers aren’t perfect, but they are exciting

My business bought two 12th gen Frameworks. First one came with a half dozen different issues that should have been picked up during QC and support was pretty lackluster. We sent that back. The second one I've been using for about six months now and am reasonably happy with. Battery life could be better (under Ubuntu) and there's way too much wobble on the screen which is exacerbated by it being glossy.

On balance though, I would probably give them another go when this needs an upgrade.

byefruiton RWKV RNN: Better than ChatGPT?

You're seeing this because the model isn't instruction fine-tuned. You'll need prompting similar to the original GPT3 or Llama models.

byefruiton Hetzner launches three new dedicated servers

Right but AWS gets to make that choice, you have no idea whether they consider a failing fan as "degradation".

byefruiton Hetzner launches three new dedicated servers

That's just the compute though, not including bandwidth or disks. These come with effectively the equivalent of ~$3k worth of AWS data transfer and vastly more SSD storage.

byefruiton Hetzner launches three new dedicated servers

Is this any worse than not even knowing when parts of your AWS instance are failing?

byefruiton Framework Laptops Now Offering Refurbished Products

Can't really speak to Fedora but Ubuntu 22.04 works almost perfectly on the 12th Gen (there's a driver you have to blacklist to get the media keys to work).

byefruiton Framework Laptop 2.5Gbps Ethernet Expansion Card

I actually think this is resonance they talk about in the last section rather than the hinges themselves. The angle at the bottom of the screen doesn't change, it's the top that flexes back and forward. Coupled with the reflective screen it makes focusing on the monitor really hard when it's on a mount to be at eye height (my Thinkpad and a Dell have no problems on the same mount).

byefruiton Framework Laptop 2.5Gbps Ethernet Expansion Card

That's fine (and why I wanted to support Framework in the first place) but do you really expect to buy a new laptop and then have to purchase more parts just to make it work correctly?

byefruiton Framework Laptop 2.5Gbps Ethernet Expansion Card

A bit sad this is arriving just as I'm getting ready to return my 12th gen Framework.

Really wanted to keep it but it arrived with a litany of quality control issues. Most of those are probably fixable with replacement parts but I don't think there's anything that can be done about the screen wobble.

A laptop I can't actually use on my lap..

byefruiton Reference count, don't garbage collect

This article provides very little evidence for it's claims and seems to only have a superficial understanding of modern GCs.

"Increments and decrements happen once and at a predictable time. The GC is running all the time and traversing the universe of GC objects. Probably with bad locality, polluting the cache, etc."

This is only the case with a mark-sweep collector, usually most of your allocations die young in the nursery. With reference counting you pay the counting cost for everything.

"In object-oriented languages, where you can literally have a pointer to something, you simply mark a reference as a weak reference if it might create a cycle."

As someone who has tried to identify memory leaks in production where someone has forgotten to "simply" mark a reference in some deep object graph as weak, this is naive.

"With an 8-byte counter you will never overflow. So...you know...just expand up to 8-bytes as needed? Usually you can get by with a few bits."

So now my "about as minimal as you can get short of nothing at all" check as an unpredictable branch in it?

"If you must overflow, e.g., you cannot afford an 8-byte counter and you need to overflow a 4-byte counter with billions of references, if you can copy it, you create a shallow copy."

I don't even know where to begin with this.

"If GC is so good, why wouldn't Python just garbage collect everything, which they already did once and could trivially do, instead of going through the hassle of implementing reference counting for everything but the one case I mentioned?"

This probably has more to do with finalising resources and deterministic destruction than anything else.

Anyone who is interested in actually studying this area would probably find https://courses.cs.washington.edu/courses/cse590p/05au/p50-b... interesting. Also https://gchandbook.org/

byefruiton Lenovo shipping new laptops that only boot Windows by default

Have had four generations of X-series Thinkpads as my daily drivers. Lenovo's build quality and customer service has deteriorated significantly over the years. Now have a 12th gen Framework on order.

byefruiton Bill to grant crypto firms access to Federal Reserve alarms experts

Genuinely interested in what products you've developed that weren't possible pre-crypto.