Hacker News Reader

who says they aren't benchmaxxing?

blackcat201on Everyone is building LLM routers, we deprecated ours

I think the author need to rethink out of the box what's LLM routers in a traditional sense ( input in, route, output ) and move to think how a router would work in agentic workflow. See cognition Devin Fusion design.

blackcat201on Gödel Agent: A self-referential agent framework for recursive self-improvement

Shameless plug, for anyone who's interested in "self-improvement" agent check out StreamBench[1] where we benchmark and try out what's essential for improvements in online settings. Basically we find feedback signal is vital and the stronger the signal the more improvement you can get if you were able to feed it back to the agent in terms of weights (LoRA) or in-context examples.

[1] https://arxiv.org/abs/2406.08747

blackcat201on Structured Outputs in the API

Do beware on some reasoning task, our recent work[0] actually found it may cause some performance degradation as well as possible reasoning weakening in JSON. I really hope they fix this in the latest GPT-4o version.

[0] https://arxiv.org/abs/2408.02442

blackcat201on 7.4 earthquake in Taiwan, 34km depth

The standard operation is to stop and check if any machine was out of calibration. So yes

blackcat201on Why host your own LLM?

I own my LLM not because I need it now but having the luxury to fall back if openai ran out of money

blackcat201on Do we really need a specialized vector database?

I have been following the vector database trend back in 2020 and I ended up with the conclusion: vector search features are a nice to have features which adds more value on existing database (postgres) or text search services (elasticsearch) than using an entirely new framework full of hidden bugs. You could get way higher speedup when you are using the right embedding models and encoding way than just using the vector database with the best underlying optimization. And the bonus side is that you are using a stack which was battle tested (postgres, elasticsearch) vs new kids (pinecone, milvus ... )

blackcat201on Ask HN: Could you share your personal blog here?

https://theblackcat102.github.io/

Recently I am ranting the AI trends and some short writeup of things I read

blackcat201on Show HN: Cerelyze: Implement research papers into code 100x faster(Desktop only)

This looks pretty interesting! But the landing page has only one sentence : Understand and implement research papers faster, followed by a get in touch button. Care to extrapolate more? The blog button doesn't work as well?

blackcat201on The company behind Stable Diffusion appears to be at risk of going under

Note that stability have been funding freelance researcher by providing compute resource such as RWKV[1], Open Assistant, some works by LAION[2] and lucidrains[3]

[1] https://github.com/BlinkDL/RWKV-LM

[2] https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K

[3] https://github.com/lucidrains/gigagan-pytorch#appreciation

blackcat201on OpenAI is now everything it promised not to be: closed-source and for-profit

On the other hand, meta AI research should be renamed as OpenAI. They are the only few big institute who open almost every models they train (galactica, OPT, M2M, wav2vec ...)

blackcat201on Underappreciated challenges with Python packaging

Recently just migrate a project from pypoetry away to the traditional setup method. Poetry works great for simple package, but once you started to add in complexities, it just falls apart due to everything was abstract away and simplified into config files and command line.

blackcat201on Ask HN: What are you working on this year?

When chatGPT first came out, my first thought was to replicate it myself. But then, I have too many missing skills or lack the time for backend, frontend and deployment. So I found LAION started an initiative for open-assistant. https://github.com/LAION-AI/Open-Assistant

So I just join them, so far its quite an active community working on pushing out the initial 0.1 version.

blackcat201on Why train when you can optimize?

You can use sin as activation function, but that would require careful initialization to avoid gradient explosion as you would ended up with a lot of points where gradient is simply zero. You can refer to Implicit Neural Representations with Periodic Activation Functions for more details.

blackcat201on Things you notice when you quit the news (2016)

What I ended up doing was crawling news from major news services and do aggregation myself. My algorithm was simple : cluster news that was reported by multple different services or multiple reports within 36 hours. By showing multiple news titles in a same box and order them by cluster size, I could just skipped those which isn't important at all.

If anyone was interested the website is available here : https://todayheadlines.live/?locale=en-US

blackcat201on Find anything fast with Google's vector search technology

I agree having a good vector is important to start with. However this is not very hard to make it work, you only need to finetune some of the clip models[1] to run it well.

Disclose: I have built a vector search engine to proof this idea[2]

[1]: https://github.com/openai/CLIP

[2]: https://unisearch.cc/search?q=how+to+reset+mac+ssd

blackcat201on The British Library puts 1M newspaper pages online for free

Do anyone know any existing effort on converting these scanned image to text corpus ( probably a new OCR model needed to be developed on these old text ) ? I think it would be more usable if they are in text form in terms of search and research purpose.

blackcat201on Mozilla Common Voice Adds 16 New Languages and 4,600 New Hours of Speech

I created edgedict [0] a year ago part of my side projects. At that time this is the only open source STT with streaming capabilities. If anyone is interested the pretrained weights for english and chinese is available.

[0] https://github.com/theblackcat102/edgedict

blackcat201on Comma Three [video]

I am surprised to see Comma Two in a consumer report not to mention out perform all other automative company including Tesla (probably the older version?). I remember in one of Geohot twitch stream bragging about their driver monitoring being really hard to duped.

blackcat201on Google now shows why it ranked a specific search result

Same in chinese search results for programming related question, most of the time I get was some bad translation from stackoverflow or github issue. Now I had to rely on browser plugins to automatically remove these domains from search results.

blackcat201on China unveils 600 kph maglev train

For reference, Boeing 747 average speed is 909 km/h (565 mph) with a top speed 986 km/h (613 mph). However maglev train saves time on landing and coordination makes it faster compare to air travel.

blackcat201on 25 Gigabit Linux internet router PC build

If anyone is interested in building 10Gb router on a budget you can buy Mellanox OCP NIC with an adapter from OCP to PCIe for around 5~30 USD ( depend on your location )[0]. I recently build one 10G router with 4x10Gb, 2x1Gb ports for around 200 USD using second hand PC.

[0] https://www.youtube.com/watch?v=hkf3PkuKdOA

blackcat201on Wav2vec Overview: Semi and Unsupervised Speech Recognition

In terms of setup and ease of use jaided OCR[1] is the best one I have used out there. I can't stress how many times I tried other alternative (tesseract, paddle OCR) out there and kept going back to use their OCR library. Mainly because it supports a large variety of languages out of the box while provide nice results.

Disclaimer : I am not affiliated to jaidedAI, just a satisfied user.

[1] https://github.com/JaidedAI/EasyOCR

blackcat201on Hacker deleted all of NewsBlur’s Mongo data and is now holding the data hostage

I am not familiar with MongoDB but is there a mechanism in which white/blacklist base on IP, just like pg_hba in Postgresql which blocks/allows only certain IP access? Few years ago when I am still using MongoDB there's only basic authentication method(user password) which blocks unwanted access. I wonder if there's anything new now

blackcat201on Brave Search beta

Any idea why this reply is downvoted? It seems there's a lot of comments that has a slight negative or doubt are downvoted in this thread. You just don't see this behaviour on any other HN thread.

blackcat201on Brave Search beta

If Brave is just another ad company (quote in their blog post) how is this different than Google?

blackcat201on AMD vs. Intel CPU Market Share

Kinda surprised on the server chart, consider most of my Linode and Vultr instance are running on AMD epyc CPUs for a years for now (at least since I last checked). I believe the market shares for AMD is much higher than the chart shows consider how low the cost per core is for a CPU slot.

blackcat201on The High Cost of Keyword Search

I recently make a vector search engine demo[0] that uses neural networks (CLIP) to encode text and image into vector. It runs surprisingly fast consider its hosted on a 2 core virtual cloud running both the text encoder and doing the vector similarity search over a 500k index results for both image and website links.

[0] https://unisearch.cloud

blackcat201on Show HN: Unisearch – vector based search engine that search across languages

It's hard to describe what's unisearch in one sentence so I will try to elaborate more in this comment.

Unisearch is basically a CLIP[0] like model that extends to multilingual domain. I also added text-text pairs to support text search for experimental purpose. So you should be able to search text or images using another foreign language.

You may try "如何重置 Mac os 系統" and "wipe format mac os" on regular website search to see the result for yourself.

I am still indexing more websites and images as of my writings. So results may improve from time to time ( since it retrieve the nearest results it can find from the database, so some answer maybe funny because it can't find a better index item )

The interface is not polish so a new image search will require you to go to the landing page and select a new image. (as this is only a final project for one of my course)

List of languages that it "supports" are English, French, German, Italian, Russian, Malay, Vietnamese, Korean, Japanese and Chinese. However I find using other language also works as well, for example text to image search using "escena nocturna al lado del río"[1] also works pretty well.

[0] https://openai.com/blog/clip/

[1] https://unisearch.cloud/search-image?q=escena+nocturna+al+la...

blackcat201on An Introduction to Knowledge Graphs

I wouldn't recommend using this work to generate knowledge graphs, as it requires a lot of rule based filtering. I suspect it wouldn't be any better than generating from a constituency parse tree. Disclaimer, I implemented this work a while ago and reach my current conclusion, still waiting for the official code to release.