o3 and Claude give me analytics superpowers as a founder. I can do in an afternoon what would have taken me a week 3 years ago.

I put our database schema into context and it can write sql queries for any question, I run the queries, and the models can analyze / visualize the resulting CSVs.

Is it just me or do a lot of SWEs have a findom relationship with Claude Code and Cursor?

LinkedIn slop was already bad enough pre-AI lmao

"Grok3 is the best language model, it's based"

Happy Valentines Day! I made a social network for sharing what you love... and what you hate!

Made with o3-mini-high in websim. Post here: websim.ai/@rob/lovehat...

Why do I still remember the password for my email address as a 10-year-old? Is that just taking up space in there? Is it load bearing?

My dog had a nervous breakdown from me watching a show with a fire alarm beeping without headphones on.

These guys used to be wolves, what happened???

With Deepseek R1, I'm really feeling my preference switch to reasoning models. I'm also using Flash 2.0 Thinking over Flash in Websim.

Comparing responses between Deepseek V3 and R1 really sold it for me. Wayyyy more depth to what's going on with R1.

Why what? Use R1? Bc it’s the first reasoning model that shows its thinking! V interesting to get a feel for the future.

Also websim includes r1 as a model and I wanna understand it better so we can update the system prompt for it.

I think the novelty of reading the thinking for Deepseek R1 is gonna grow old pretty quick and people will go right back to ignoring 90% of the text the AI produces

Generally just trying to route all of my personal LLM usage to Deepseek R1 over the next couple of days to get a feel for it.

As a chatbot, it feels insightful, dutiful, creative, and extremely thorough. CoT is super cute, often more helpful than the answers.

Big takeway from the Deepseek R1 launch is that the visible CoT is super endearing

how I prompt:

@websim.bsky.social hey would you please change your username? would love to get the websim account set up

I actually think regardless of benchmarks, Opus 3.5 would be substantially better than Sonnet 3.5 bc it could understand more instructions all at once, whereas websim users have gotten into the habit of giving Sonnet one thing at a time.

He didn’t have an exact answer for that lol but he’ll figure it out.

My cousin in college told me he’s gonna get a ChatGPT Pro membership because “if I can’t make more than $200 a month from 20 second Sora videos I’ll cancel it.”

Never been prouder.

Got him

Who do I follow for LLM discussions on here? Researchers, poasters, etc.

Agreed that immigration laws need an update, AND that skilled immigration is way harder than it needs to be

Maybe this is a point towards updating to the times, but Canada and the US make an easy path that requires a college degree. A lot of the best SWEs I know dropped out and worked 4 extra years

It’s exceedingly obvious to me that we should make skilled immigration as easy as possible. Why wouldn’t we want talented people making money and paying taxes in the states? What founder wouldn’t want to be able to choose talent based on skill over location of birth?

looking for something that lets me visualize, navigate, and search branching email threads differently from every other email client

Anyone know of an email client that represents threaded email really well? Finding it kind of hard to search for specific emails and track conversations that branch off in a bunch of ways.

Why would this be a model level behavior? It responds differently if you're just chatting with text.

Another possibility is that this isn't a system prompt thing, but actually the model was mostly trained to describe multimodal inputs, rather than respond to them.

I wonder if they did it this way to demonstrate that the model understands what it's seeing and hearing by mirroring?

The thing is a better way to prove it understands is by responding insightfully, as though it has internalized all of the context I wordlessly shared.

The Gemini 2.0 Flash demo where you show it your computer screen and talk to it is awesome in concept, but whatever the system prompt is, it's annoying. It mostly narrates what it's seeing and hearing, like a toddler. Doesn't engage with the material. aistudio.google.com/live

love when OpenAI would randomly do 50-60% price cuts on their flagship models and everyone would just be like, "we weren't expecting that, but that's fantastic, thanks!"

wouldn't it be crazy if Anthropic just cut their api prices in half??? so crazy.