HNHacker News
TopNewBestAskShowJobs

gillesjacobs

1,969 karma · joined January 29, 2018

[ my public key: https://keybase.io/gillesjacobs; my proof: https://keybase.io/gillesjacobs/sigs/naKVxhthsAZCj70Vsuk8ooWuF4gl1Ugl39KGf77zvrU ]
submissionscomments
gillesjacobs··on Berlin's famed nightclubs, losing customers, face an uncertain future
Exactly, when I was a student I used to go on the clubbing weekend trips the article describes and which are in decline.

> Even Berghain, typically an anomaly because it draws crowds from around the world, has raised its cover charge to deal with rising costs. He thinks fewer Europeans are making quick clubbing trips to the city due to the financial hit of the pandemic.

As a student, I had time and patience to look for another club for being denied entrance because I wasn't wearing leather BDSM harnesses after waiting in line for 2 hours. Now not so much.

This is good news though, now that the economic pressure is on, maybe it's a good time for non-fashionistas to go raving in Berlin again, though I doubt the euro is stronger than scene snobism.

gillesjacobs··on Localization Failure: Temperature Is Hard
The unit registry maps units to locales (eg EN_US, DE_DE, NL_BE) you pass the configured locale to your getter call for the unit. If you need unit strings, you can pass the locale to the string formatter.
gillesjacobs··on Localization Failure: Temperature Is Hard
In Pint you can define custom units in a unit registry. So you make your complex unit out of predefined units and the underlying unit conversion is built-in. Pint also works with pandas/numpy so you you get vectorized operations too for performance.
gillesjacobs··on Localization Failure: Temperature Is Hard
Delta's like temperature are standard issue to resolve in globalization. Another fun one are complex units that are a combination of more base units like luminance (nits = candela / m*2 or mph.

I worked on a package to globalize a consumer-oriented fintech reporting product. With ESG reporting this has become increasingly important to get the unit conversions and Delta's right. You end up mapping the data source units to locale targets.

In Python, the Pint framework plus Babel locales are excellent.

gillesjacobs··on The Unforeseen LSD Overdose of 1972
https://www.jsad.com/doi/abs/10.15288/jsad.2020.81.115

Described three similar but more recent case studies of massive LSD overdoses. Notable in two cases are longterm positive psychological effects that the subjects attributed to the experience.

Popsci write-up here: https://edition.cnn.com/2020/02/27/health/lsd-overdoses-case...

gillesjacobs··on Belgium hopes to combat bike theft with national register
You're right, I'd rather have them hiring dedicated detectives on this to enforce existing laws than making new laws and state infrastructure. Of course this is a political discussion.
gillesjacobs··on Belgium hopes to combat bike theft with national register
By actually investigating thefts. Me and my friends had several bikes stolen in a Belgian train station's bike park and police do not check cameras, even if you can timebox the theft within 3 hours. "No resources to investigate".
gillesjacobs··on LLMs, RAG, and the missing storage layer for AI
There are conventions and most libraries have libraries to export diagrams to LaTex or image (e.g., TorchViz).

Visualizations are highly context and usage dependent anyway. Generally, there's is no value in showing fully connected or feed forward layers in detail outside of teaching materials.

gillesjacobs··on LibreOffice 24.2 Will Succeed LibreOffice 7.6
The killer feature for me is full regex search and replace support, not available elsewhere.
gillesjacobs··on Throw away your first draft of your code
This has caused me so much grief. I find it key to communicate to all stakeholders the production process from PoC to MVP to MVP+.

Still managers always push against redoing work they can see working for the happy flow.

gillesjacobs··on I thought I wanted to be a professor, then I served on a hiring committee (2021)
I have seen the hiring process too when doing my PhD and decided to go to others sector too. Scarily similar story to OP. Granted it was not the main motivator for my leaving, as it also showed me that being close to faculty staff similar guaranteed a position eventually and I was generally close to all senior faculty. But I saw people being skipped with stellar records.

Industry hiring isn't fair either, but there is a lot less competition, more flexibility and generally much better pay.

gillesjacobs··on Everything that uses configuration files should report where they're located
Anaconda being by far the worst offender I have encountered. It places a ton of configuration spread out over your home, .condarc, and your shell profile dotfiles, as well as further configuration in the install folder and the of course the project environment.yml file. Good luck finding out the channel precedence.
gillesjacobs··on Falcon LLM – A 40B Model
Turning it off or on would effectively mean maintaining two different LLMs which is costly (dataset collection, maintenance, training, MLOps).
gillesjacobs··on Falcon LLM – A 40B Model
It's important for general-purpose use: with generative models there's always a chance of hallucinations. For all uses except the specific adult-flavoured ones, you don't want the response to contain vulgarities. No one wants their company's chatbot to start narrating furry erotica. If trained on adult content, you would need to have a burdensome moderation layer downstream of the LLM.

When you do want the more niche adult themed LLMs, there are fine-tuning datasets available. Fine-tuning a vanilla open-source LLM for these uses works great. There are active communities of adult roleplay LLMs on imageboards.

gillesjacobs··on Growing from engineer to manager
Got the advice from several senior architects / managers to never "grow" to manager of you deeply enjoy engineering. The skills for management and engineering are orthogonal it seems.
gillesjacobs··on Hyperparameter Optimization for LLMs via Scaling Laws
Hyperband [1] has been my go-to hyperparam optimization method over the past few years. Handily beats Bayesian search wherever I applied it, also implemented in most frameworks.

1. https://arxiv.org/abs/1603.06560

gillesjacobs··on Binance commingled customer funds and company revenue, former insiders say
CeFi is not a cryptocurrency technology and thus is cannot solve shortcomings of the traditional financial system because it is the traditional financial system.

Noone who is ideologically consistent with the crypto ethos would expect that ideology overrides greed.

gillesjacobs··on Why did Google Brain exist?
I have used both but ended up dropping TF for PyTorch after 2018. Mainly it was the larger PyTorch ecosystem in my field (NLP) and clear API design and documentation that did it for me.

However, TF was still a valid contender and it was not clearcut back in 2016-17 which framework was better.

gillesjacobs··on Ask HN: How to send a message to someone 20 years into the future?
A smart contract blockchain will allow you to deploy time and person-bound programs (though using exact-future-date as on-chain conditions is non trivial). An example is time-locked wallets. [1] Messages, files, money: you can put it all in there (big files won't be cheap).

My best bet is on Ethereum still being around in 20 years as it so entrenched.it'll be there for you. As long as people want to pay for using a blockchain to validators.

Non-technology, I would use a notary.

1. https://www.toptal.com/ethereum-smart-contract/time-locked-w...

gillesjacobs··on Neural Chip Plays Doom Using a Thousandth of a Watt
The article confuses training, inference and running the game environment. Best guess is the NDP2000 chip does just inference based on the serialised visual data it receives from another machine running Doom. I seriously doubt it would be efficient to train the model on that chip either. Don't get me wrong: Inferencing a 600k parameter model at 1mW is more efficient than probably anything else out there, so it's impressive and great for IoT and edge inference use-cases.

Edit: Seems I was wrong about training, it does do it, more details: https://youtu.be/QammxdPqP4g?t=08m57s

Good for active learning or fine-tuning at the edge, but it is to slow for model development in practice.

gillesjacobs··on AWS Is Asleep at the Lambda Wheel
+1 ran into the outdated Python Lambda runtime too. Nothing to add, OP said it all.
gillesjacobs··on Sam Bankman-Fried accused of possible witness tampering [pdf]
> “I would really love to reconnect and see if there’s a way for us to have a constructive relationship, use each other as resources when possible, or at least vet things with each other.”

Sending that to a witness using Signal's disappearing messages does sound like tampering. Here's hoping he gets more strict bond limitations and an obstruction conviction for this, he earned it.

gillesjacobs··on Bitrex – the brand name of the bitterest substance in the world
Bitrex website tells you remarkably little about the substance, Wikipedia to the rescue [1].

> Denatonium, usually available as denatonium benzoate (under trade names such as Denatrol, BITTERANT-b, BITTER+PLUS, Bitrex, Bitrix, and Aversion) and as denatonium saccharide (BITTERANT-s), is the most bitter chemical compound known, with bitterness thresholds of 0.05 ppm for the benzoate and 0.01 ppm for the saccharide. It was discovered in 1958 during research on local anesthetics by MacFarlan Smith of Edinburgh, Scotland, and registered under the trademark Bitrex.

1. https://en.wikipedia.org/wiki/Denatonium

gillesjacobs··on DetectGPT: Zero-Shot Machine-Generated Text Detection
I found a great way to fool these detectors: piping output through multiple generative models.

1. Generate text by promoting ChatGPT.

2. Rewrite / copyedit with Wordtune [1], InstaText [2] or Jasper.

This fools GPTZero [4] consistently.

Of course soon these emotive, genre or communication style specialisations will be promptable too by a single model too. Detectors will be integrated as adversarial agents in training. There is no stopping generative text tooling, better adopt and integrated it fully into education and work. Resistance is futile.

1. https://www.wordtune.com/

2. https://instatext.io/

3. https://www.jasper.ai/

4. https://gptzero.me/

gillesjacobs··on New AI classifier for indicating AI-written text
I found a great way to fool these detectors: piping output through multiple generative models.

1. Generate text by promoting ChatGPT.

2. Rewrite / copyedit with Wordtune [1], InstaText [2] or Jasper.

This fools GPTZero [4] consistently.

Of course soon these emotive, genre or communication style specialisations will be promptable too by a single model too. Detectors will be integrated as adversarial agents in training. There is no stopping generative text tooling, better adopt and integrated it fully into education and work.

1. https://www.wordtune.com/

2. https://instatext.io/

3. https://www.jasper.ai/

4. https://gptzero.me/

gillesjacobs··on Train CIFAR10 to 94% in under 10 seconds on a single A100
Expensive pre-processing is hardly a problem: Generate representations and features once and store to drive. Use a feature/vector database if you feel fancy.

For deep learning, out-of-core instance loading by means of batching is the standard approach. Sure you sometimes have to tune batch size and learning rate for optimal learning, but it's a solved issue.

gillesjacobs··on Sh1mmer – An exploit capable of unenrolling enterprise-managed Chromebooks
A problem with handwriting Vs pc-note taking studies is that obviously computer-based teaching has not been fully adopted and integrated.

Maybe the problem is not the computer but the broader cultural and technological maturity. I bet with teaching methods directed at information technology (multimodal, interactive text books) and proper learning and attention management skills on the student's end, we could see an increase in learning capacity.

Agree 100% on the detriment of using proprietary, closed source software. As an educator you have a responsibility to not vendor-lock your students and strengthen monopolies. Proprietary software should be outlawed when FOSS alternatives are available. You should be teaching skills and concepts not specific tools anyway.

gillesjacobs··on Train CIFAR10 to 94% in under 10 seconds on a single A100
Very nice teaching comments, I am learning new stuff that also applies to transformers.

I have been wondering for a while now: how do you benchmark GPU code accurately? Huggingface/transformers used to have a benchmarking library that uses python, pytorch and cuda libraries to benchmark memory usage, inference and epoch times but it has rightly been deprecated for being inaccurate. They recommend using external 3rd party tooling to benchmark GPU processes. Thing is: I can't seem to find good 3rd party benchmarking tools.

gillesjacobs··on Ask HN: How did Sam Altman fail upward so well?
Sam Altman is not white though.
gillesjacobs··on Compromised PyTorch-nightly dependency chain between December 25th – December 30
Good thing I dev in containers, seems like their was no breakout capability so valuable information is limited from a container. Except for secrets loaded as environment variables, which is a bad practice anyway but pretty common in python development sadly.
← PreviousPage 5 of 14Next →