Iwona Bialynicka-Birula ⏩

The em dash has become the de facto watermark slapped onto AI-generated text... and onto the people who are not tech-savvy enough to remove the watermark before trying to pass AI generations as their own writing

Next-level vibecoding requires next-level context-switching skills: vibecoding at full speed while spending the time during which the assistant is coding on doing other things

Good old 4.5 and I are on the same page w.r.t. alignment

It's amazing how often people include conflicting instructions in their megaprompt and then scratch their heads wondering why the LLM doesn't adhere to them

They version the models in the most confusing ways and then they wonder why the models think 9.11 > 9.9

reverse Ghiblification is way harder than regular Ghiblification

comment if you can do better than this

Even the icons have bad posture

Google's new AI Mode speed is rivaled only by its level of inaccuracy. How do you even somehow get worse at the strawberry question after being given access to all the articles on the internet which talk about the very issue?

Don't get me wrong, I enjoy your Ghiblified family photos as much as the next guy, but didn't we already solve style transfer like 10 years ago?

💡System Design Interview, but you also have to vibe-code the entire working system within the span of the interview.

No, Apple did not shoot itself in the foot by betting on privacy at the cost of AI. It outright kneecapped itself.

I'm finding it harder to vibe-code after alcohol than to regular-code, which is the ultimate proof that it's a harder cognitive task (makes sense - all the easy parts were displaced by LLMs)

AI coding assistant gods,

grant me the serenity to accept the changes which are correct,

courage to reject the changes which are not,

and wisdom to know the difference

🙏

Nvidia making a cute little droid instead of another creepy uncanny valley humanoid makes it the only big tech company with a semblance of emotional intelligence

found a note I left for myself
don't remember writing the note
assume I'm an LLM

What if the "how is babby formed?" question author was on to something? Phrasing the question in this way certainly helps get around OpenAI guardrails...

OMG, were all these people asking Yahoo dumb questions from the future, trained to prompt LLMs to bypass guardrails? 🤯

New rule: everybody who says "PhD-level" has to specify what the PhD is in. I mean, is it aerospace engineering or gender studies? There's got to be a difference.

Google Maps: ETA 5:21
I decide to accelerate
I decide to go 80 instead of 70 the entire way
I arrive at exactly 5:21
Nothing like Google Maps to remind you that free will is an illusion

Switched from claude-3.5-sonnet to claude-3.7-sonnet in Cursor and 3.7's like "I don't know who you had here before me, but boy did they do a shoddy job, let me rewrite all this code for you my way..."

The p(doom) of the bots mirrors that of their creators:

Microsoft Copilot: 0%
Meta AI: 0%
Google Gemini: 0%
xAI Grok: 10%
OpenAI ChatGPT: 15%
Anthropic Claude: 30%

The sheer amount of tech debt you can accumulate in just one weekend of vibe coding starting from scratch is... staggering

DeepSeek-R1-Zero prompt

I sometimes wish people would also "think about the reasoning process in the mind" before answering.

NotebookLM: you know you found product-market fit when SNL does a skit about you

DeepSeek's thought process in response to "Hi": Okay, the user said "Hi". That's pretty straightforward. I should respond in a friendly and welcoming manner. Maybe say hello and ask how I can assist them today. Keep it open-ended to encourage them to share what they need help with. Let me check the previous interactions to make sure there's no specific context I'm missing. Nope, looks like this is the start of the conversation. Alright, time to draft a response that's warm and inviting. Let me make sure the tone is right. Not too formal, but still professional. Something like, "Hello! How can I assist you today?" That should work. Wait, maybe add a smiley emoji to make it friendlier? Hmm, the guidelines don't mention anything against emojis, but maybe keep it simple. Stick with the basics. Yep, that's good. Ready to send.

What happens in your coworker's mind when you send them a lone "Hi" on Slack

"The handshake in space"

It warms my heart how despite all of the US vs. China "adversary" rhetoric, engineers tend to ignore the divisiveness, give credit where credit is due and celebrate each other's achievements regardless of borders. It reminds me of this beautiful iconic moment in human history.

If this is Meta, I can only imagine what is going on at Google

DeepSeek's answer to "What is 2+2?" is long and full of convoluted reasoning and second guesses. Towards the end it mentions "Maybe the user is testing if I overcomplicate things."

Yes, DeepSeek, you do overcomplicate things 😂

Cartoon of cavemen inventing the wheel with one saying "ONCE THIS THING CATCHES ON PEOPLE WILL NO LONGER HAVE JOBS"

This is how I think some of the discourse will seem in hindsight

cointelligence 💯

"Dependency" XKCD (https://xkcd.com/2347/) with "ALL MODERN DIGITAL INFRASTRUCTURE" replaced by "AI Bluesky" and "A PROJECT SOME RANDOM PERSON..." replaced with "Nathan Lambert"

Just kidding :)