agi a

[derek guy @dieworkwear]

DRESSING FOR LARGER MALE FIGURES 🧵

Want to do a quick thread on how to dress if you're a larger guy. There's not a lot of info or inspiration online about this, and I feel this leads people to think that they don't have options or can't dress well bc of their body type.

[LLM Security @llm_sec]

Low-Resource Languages Jailbreak GPT-4

paper: https://t.co/B457h1LcNj

"Our work exposes the inherent cross-lingual vulnerability of these safety mechanisms, resulting from the linguistic inequality of safety training data, by successfully circumventing GPT-4's safeguard through translating unsafe English inputs into low-resource languages. On the AdvBenchmark, GPT-4 engages with the unsafe translated inputs and provides actionable items that can get the users towards their harmful goals 79% of the time"

with @yong_zhengxin, @CriMenghini, @stevebach

[Siqi Chen @blader]

It's depressing to see so many talented people from top schools with amazing resumes go work years at FAANG, and the most they can point to having shipped are a few buttons in one minor workflow.

I mean I know FAANG can afford it, but you can't. Work at a startup!

[21st Century City @urbanthoughts11]

Excuse me, while I claim my street space
by Michael Rakowitz (2004)

[Mark Tenenholtz @marktenenholtz]

3 months ago it was commonly accepted that OpenAI had an insurmountable lead.

Since then:

• 65B models training on 1 GPU
• Llama 1 + 2 with a massive ecosystem built around them
• Efficient, open-source C inference code
• Instruction-tuned models (and data) are plentiful

A handful of us have seen the trajectory this whole time. Try to find more second order thinkers.

[Akiyoshi Kitaoka @AkiyoshiKitaoka]

Chromostereopsis

The majority of people see red in front of blue, while 10-20% see blue in front of red.

[David Senra @FoundersPodcast]

Steve Jobs: Asking for help is a superpower

“I’ve never found anybody that didn’t want to help me if I asked them for help.

I called up Bill Hewlett (founder of HP) when I was 12 years old.

He answered the phone himself.

I told him I wanted to build a frequency counter. I asked if he had any spare parts I could have.

He laughed.

He gave me the parts. And he gave me a summer job at HP working on the assembly line putting together frequency counters.

I have never found anyone who said no, or hung up the phone.

I just ask.

Most people never pick up the phone and call. And that is what separates the people who do things, versus the people who just dream about them.

You have to act.”

[Sebastian Raschka @rasbt]

If you're looking for an nice read over the long weekend, I highly recommend the recently uploaded survey "Instruction Tuning for Large Language Models" by Zhang et al.: https://t.co/QqZkNikWN1

Covers both the creation and usage of instruction-datasets for LLM finetuning.

[Peter Yang @petergyang]

A personal story about Satya Nadella's superpower from a Microsoft VP:

"A few weeks after I joined Microsoft, Satya randomly called me and my manager to chat.

During those 30 minutes, he only asked questions. He asked about our thoughts on the product strategy, Microsoft’s culture, and what we thought needed change.

Here you have this CEO of a $2 trillion company just listening to two new employees instead of telling us what to do. That’s remarkable.

Later on, I realized that this is how Satya gathers signal.

He’s really good at getting different points of view from different sources and then connecting the dots on what needs to be fixed.

It’s remarkable that he can do this without falling into the temptation of telling you what to do."

As Satya himself wrote in his book, Hit Refresh:

"Listening is the most important thing that I accomplished each day because it would build the foundation of my leadership for years to come."

[Jikku Jose @jikkujose]

Just cancelled my ChatGPT subscription (after subscribing from when it came out).

Found a great self hosted alternative that lets me chose the precise model that I want.

Thinking of providing this for a set of early users (~20). Retweet if you want an access code.

[Eade @eade_bengard]

get in your car, drive to your apartment, read some of the book your friend recommended, scroll Instagram before bed

[Jim Fan @DrJimFan]

In short, memory bandwidth is all you need to run “attention is all you need”.

[agi a @AntonAlpen]

@culturaltutor So this is how pixels and HDR were invented.

[derek guy @dieworkwear]

i saw a therapist last yr bc i kept feeling like everyone hates me, and my therapist kept saying, "just bc you feel something doesn't mean it's a fact," and about a year later, NBC has a headline story that says "everyone hates this guy"

https://t.co/AeV4AxFdju

[Arvind Narayanan @random_walker]

Prompt injection keeps getting wilder. Unlike text-based methods, you can't tell by looking at one of these images or audio clips that it's malicious. It can be embedded in a website/email attachment and will alter the model's behavior if it processes it.
https://t.co/NbzgKyZFv6

[Arvind Narayanan @random_walker]

The "autoregressive trap" is when a chatbot commits to an answer and then says whatever it takes to justify it. You can reliably trick it by giving it a problem that requires reasoning — but ask it to output the answer first.

HT @goodside who coined the term.

[Science Is Strategic @scienceisstrat1]

The US 🇺🇸 economy and the EU 🇪🇺 were roughly the same size in 2008.

The US is now nearly twice as big 📊

[murat 🍥 @mayfer]

@goodside this just means the data sets used aren't good for testing embedding usefulness i.e match texts that diverge on wording but converge on meaning

the over-trust of haphazard test sets is fueling an incredible amount of false hype in the space

[Riley Goodside @goodside]

this is wild — kNN using a gzip-based distance metric outperforms BERT and other neural methods for OOD sentence classification

intuition: 2 texts similar if cat-ing one to the other barely increases gzip size

no training, no tuning, no params — this is the entire algorithm:

[Kristen Anderson @FintechKristen]

Public service announcement: two children in daycare at @BrightHorizons in Cambridge, MA costs $95,400/year.

This is after-tax money (ie about $130k in income would be needed to afford this).

Shame on this country.

Cc @reshmasaujani