Hacker News Reader

m3aton Why I'm still bearish on LLMs after Navier-Stokes

Yes "oversight and guardrails" are still needed, but even that is becoming easier to build, and imo already no longer in the "laborious" category. Even far from the frontier, you can tune a 0.2B LLM into a decent 2000 Elo player as a weekend project:

https://x.com/maximelabonne/status/2100137121264828901

m3aton Engineers will do anything to avoid learning from history

> Working with AI feels more like leadership than coding

m3aton Granite 4.1: IBM's 8B Model Matching 32B MoE

https://research.ibm.com/blog/granite-4-1-ai-foundation-mode...

Original article on IBM research

Hugging face weights: https://huggingface.co/collections/ibm-granite/granite-41-la...

m3aton Go away Python

As shared in a sibling comment, you can get away with just curl+shell: https://paulw.tokyo/standalone-python-script-with-uv/

m3aton Go away Python

And with some small shebang trick, you don't even need to have uv installed [1], just curl and a posix shell

[1] https://paulw.tokyo/standalone-python-script-with-uv/

m3aton Accepting US car standards would risk European lives

> threatening to start a new trade war if the EU doesn't permit their murdermobiles on the European roads

The strange part is that those car can be sold in the EU markets already. They just have to comply with the same pollution and safety standards as other cars. What would justify an exception?

m3aton Why I love OCaml (2023)

> mostly because it's the scripting language of choice for PyTorch and AI-adjacent libraries/tooling/frameworks

I would politely disagree. Torch started in Lua, and switched to Python because of its already soaring popularity. Whatever drove Python's growth predates modern AI frameworks

m3aton Introducing tmux-rs

Very much a side note, but:

> my feeling is that I’d still reach for it if my hands are really physically hurting, and I need to keep working. Usually once I reach the point where I’ve got blisters on my fingers I think it’s better to just take a break

I'm dumbfounded, and impressed in an unhealthy way. Do some of you regularly type so much that you develop blisters?

m3aton Launch HN: Vassar Robotics (YC X25) – $219 robot arm that learns new skills

You're in luck: https://github.com/Vector-Wangel/XLeRobot

m3aton Ask HN: Who is hiring? (February 2025)

Hi Bryant, I'm quite interested in the senior software engineer role in the applied AI team!

On your job board I'm seeing slightly different infos about the location requirement. Can you clarify if the role is globally remote or US only?

m3aton Using uv as your shebang line

Uv's support of inline metadata is super handy. What if the person running the scripts doesn't have uv yet though? For fun I wrote a double shebang script that will handle that too: https://paulw.tokyo/standalone-python-script-with-uv/

m3aton Textual 1.0 Release

Accompanying blog post: https://textual.textualize.io/blog/2024/12/12/algorithms-for...

m3aton Ask HN: Who wants to be hired? (September 2024)

Marketer, with SDG specialty

Location: Europe or Japan

Remote: Flexible

Willing to relocate: Yes

Skills: international marketing professional (7 years). Decided to specialize further, I obtained a master in globalization, business and development from the University of Sussex. Speak English and Japanese, intermediate in French.

Résumé/CV: https://www.linkedin.com/in/yukie-soeda-319499196 (email me for a pdf)

Email: 901stb {@t} gmail.com

(independent from m3at's account)

m3aton Insights from over 10,000 comments on "Ask HN: Who Is Hiring" using GPT-4o

Yeah langchain is not necessary for this. The author appear not to have shared his code yet (too bad, the visualizations are nice!), but as a poor replacement I can share mine from over a year ago:

https://github.com/m3at/hn_jobs_gpt_etl

Only using the plain OpenAI api. This was on GPT-3.5, but it should be easy to move to 4o and make use of the json mode. I might try a quick update this weekend

m3aton Tokens, n-grams, and bag-of-words models (2023)

Two of the most popular libraries for token creation (tokenization) are in fact in rust, with an interface in python:

https://github.com/huggingface/tokenizers

https://github.com/openai/tiktoken

m3aton Intel discloses $7B operating loss for chip-making unit

That's a lot! More than half of the yearly buybacks Intel did in 2018-2020! /s

https://www.intc.com/stock-info/dividends-and-buybacks

m3aton Piano on Paper

I agree for pianos, because the mechanical feedback is part of what makes playing enjoyable.

I'm not as sure for keyboards in VR! There has been a lot of research on non-invasive brain computer interface (BCI), including predictive systems that guess what you want to type, instead of where you're actually typing (example from 8y ago [1]).

Simpler gesture recognition is already on the market (like this one, from 2020, for the apple watch [2]). And now bigger VR players are investing in the tech [3]. I expect useful brain interface to be integrated in common VR devices in a couple of generations.

[1] Air Keyboard: Mid-Air Text Input Using Wearable EMG Sensors and a Predictive Text Modeland a Predictive Text Model; https://digitalcommons.dartmouth.edu/cgi/viewcontent.cgi?art... [2] https://mudra-band.com/ [3] https://www.androidcentral.com/gaming/virtual-reality/zucker...

m3aton JAXA SORA-Q mini moon rover

The rover is expected to land in 5 days: https://twitter.com/SORAQ_official/status/174680914003773861...

m3aton Is Cinema Becoming Less and Less Innovative with Time?

Direct link to the pdf: https://ceur-ws.org/Vol-3558/paper7806.pdf

m3aton Gemini AI

For others that were confused by the Gemini versions: the main one being discussed is Gemini Ultra (which is claimed to beat GPT-4). The one available through Bard is Gemini Pro.

For the differences, looking at the technical report [1] on selected benchmarks, rounded score in %:

Dataset | Gemini Ultra | Gemini Pro | GPT-4

MMLU | 90 | 79 | 87

BIG-Bench-Hard | 84 | 75 | 83

HellaSwag | 88 | 85 | 95

Natural2Code | 75 | 70 | 74

WMT23 | 74 | 72 | 74

[1] https://storage.googleapis.com/deepmind-media/gemini/gemini_...

m3aton Gemini AI

I thought so too but no, Gemini _Ultra_ is supposed to beat GPT-4. The one available through Bard is Gemini Pro. The wording sure is confusing!

Looking at the technical report [1], on selected benchmarks, rounded score in %:

Dataset | Gemini Ultra | Gemini Pro | GPT-4

MMLU | 90 | 79 | 87

BIG-Bench-Hard | 84 | 75 | 83

HellaSwag | 88 | 85 | 95

Natural2Code | 75 | 70 | 74

WMT23 | 74 | 72 | 74

[1] https://storage.googleapis.com/deepmind-media/gemini/gemini_...

m3aton Unciv

I can recommend mindustry as well: https://mindustrygame.github.io/

m3aton Jina AI launches open-source 8k text embedding

This is not an embedding model though. Yes you can always extract some embeddings from somewhere, but for most LLMs those won't perform well for retrieval (which makes sense as it's not what the models are optimizing for)

m3aton Jina AI launches open-source 8k text embedding

To take an other example, would you call a game that has its code and all assets (ex. character sprites) freely available open source? Or would the process that was used to create the assets in the first place also be required to be considered open?

The parallel can be made with model weights being static assets delivered in their completed state.

(I favor the full process being released especially for scientific reproducibility, but this is an other point)

m3aton Embeddings: What they are and why they matter

I would only recommend pgvector if you're already primarily relying on postgres and the scale is limited (<1M documents). It won't handle the part that generates the embeddings though. You could use cloud vendors if you're in one of their ecosystem, do it yourself [1] (but model serving can be tricky without prior experience in ML), or use some other service to generate embeddings [2].

Alternatively vespa cloud [3] offer both but… not the easiest to work with, it's tailored for businesses where search is a primary component.

Feel free to shoot me an email (in profile) with your context if you have more questions, in case I can help

[1] This model is a solid baseline if you're working with English text: https://huggingface.co/sentence-transformers/all-mpnet-base-...

[2] OpenAI's embeddings is probably the easiest to get started, and the API is straightforward. It's not the best performing embeddings for retrieval but good enough in some cases: https://platform.openai.com/docs/guides/embeddings/use-cases

[3] https://cloud.vespa.ai/

m3aton Embeddings: What they are and why they matter

Depends what you mean by taking all your articles. If it's scraping no, but if you provide text content and urls as key/values pairs yes.

SageMaker and VertexAI are the AI services of AWS and GCP respectively, and they both offer embedding generation and vector databases (the two key pieces necessary for embedding search).

There are a bunch of smaller companies offering vector search as a service too, example pinecone to name just one: https://www.pinecone.io/

m3aton Undermining Democracy: The EU Commission's Controversial Push for Surveillance

I learned about this effect too recently in an ACX post, let me add a link for the curious: https://www.astralcodexten.com/p/your-book-review-secret-gov...

m3aton TimeGPT-1

I was surprised too!

While I could find some excuses to exclude ARIMA, notably that in practice you need to input some important priors about your time series (periodicity, refinements for turning points, etc) for it to work decently, "prohibitive compute and extensive training time" are just not applicable.

That part is a bit wanky, but the rest of the paper, notably the zero shot capability, is very interesting if confirmed. I look forward for it to be more accessible than a "contact us" api to compare to ARIMA and others myself

m3aton Ask HN: Who is hiring? (October 2023)

Thanks for sharing, your mission is highly interesting! The DS lead role could be an ideal match, though if you allow me to be blunt to make the best of both of our time: is there flexibility in the package?

If yes I'll reach out, or feel free to do so (email in profile)

m3aton Firefox address bar

Those prefix are great and I use them regularly!

Though imo the killer feature of FF address bar is simply that it's tied to a proper search history. Unlike chrome (which I sadly have to use at work), that only keep 90 days of history (!), making the address bar useless for anything but tabs, recent searches and as a link to a search engine. I really can't see an excuse for that behavior, the sqlite used by chrome is a few mb at worst.