Ben Cipollini

Founding AI Engineer @ @Galileo_AI. Former Facebook/Meta, Classy, GivingTuesday, MSFT. Using AI to solve real-world problems~

[Ben Cipollini @mrquickowl]

Spending time in Pacific Grove these days, I get to see these butterflies on their migration.

Today is #GivingParksday, a day to support CA state parks. They provide critical habitats to wildlife. Any donation to @calparks will be matched up to $50K! 🦋 https://t.co/B5BA01ZWCz

[Ben Cipollini @mrquickowl]

Excited by RouteLLM - not just using routing to achieve cheap performance, but the extensible training paradigm for training routers. This makes so much sense - can't wait to try it out!

[Arena.ai @arena]

Not all questions need GPT-4!

We introduce RouteLLM – a routing framework based on human preference data that directs simple queries to a cheaper model.

With data augmentation techniques, RouteLLM achieves cost reductions of over 85% on MT Bench and 45% on MMLU while maintaining 95% GPT-4 performance.

Comparing against commercial offerings (Martian and Unify AI) on MT Bench and achieve the same performance while being over 40% cheaper.

Our model, datasets, and code for serving and evaluating LLM routers are all open-sourced. We are excited to see what the community will build on top!

1/6 (blog post & links👇)

[Arena.ai @arena]

Not all questions need GPT-4!

We introduce RouteLLM – a routing framework based on human preference data that directs simple queries to a cheaper model.

With data augmentation techniques, RouteLLM achieves cost reductions of over 85% on MT Bench and 45% on MMLU while maintaining 95% GPT-4 performance.

Comparing against commercial offerings (Martian and Unify AI) on MT Bench and achieve the same performance while being over 40% cheaper.

Our model, datasets, and code for serving and evaluating LLM routers are all open-sourced. We are excited to see what the community will build on top!

1/6 (blog post & links👇)

[Ben Cipollini @mrquickowl]

Success indeed - nearly 70% of people using this pass to visit a state park earn <$60K/year.

Encourage everybody to sign the petition, to get the minimal state funding to continue this highly successful - and impactful - program.

[California State Parks Foundation @calparks]

We need your help! 📣 Despite the success of the California State Library Parks Pass, without new funding, libraries cannot provide new passes for checkout.

Tell your legislator to restore funding for this popular and effective park access program! 👉https://t.co/vjeAWcoMZO.

[Ben Cipollini @mrquickowl]

Excited to be a part of this product launch!

@Galileo_AI unlocks design-centered product ideation, and helps me communicate with designers more efficiently and effectively.

Can't wait to see amazing ideas get out of people's heads and into Galileo-based designs! Let's go!

[Stitch by Google @stitchbygoogle]

Meet Galileo 1.0: The game-changing prompt-to-UI platform now open to all.

Our mission is to supercharge the creative process of designers and builders, making design easy, fast and magical.

Bring your idea, let AI do the heavy lifting.

Start today: https://t.co/a1xWOnIrQC

[Ben Cipollini @mrquickowl]

Love the interplay between AI features and frontend magic 🪄🪄 Well done team!

[Stitch by Google @stitchbygoogle]

Happy New Year from Galileo AI! 🎉

We're excited to start 2024 with a new feature drop: Styling Customization 🎨.

After you generate a design through our AI model, you can customize theming options like color, font, border radius.

Light/Dark Mode: Easily switch between ☀️ and 🌘.
Color Palette: Choose from popular preset color palettes or select a primary color to automatically adjust the rest of color shades.
Corner Radius: Select from options like sm (4px), md (8px), lg (12px), or pill (fully rounded).
Font: We have a selection of commonly used fonts for UI design, including both sans serif and serif options.

Some cool examples:👇

[Ben Cipollini @mrquickowl]

Excited to be able to translate and share app designs with my 日本人 friends so easily. 🇯🇵

Some really cool innovations behind the scenes to make this happen - congrats to the @Galileo_AI team for landing it!

[Stitch by Google @stitchbygoogle]

🌐Say hello to easy localization design✨

Introduce our new multi-language UI generation. Create your design in Galileo, say the magic word, and boom - your text flips to French, German, Chinese, Japanese, Hindi, or any other languages you’d like.

[Ben Cipollini @mrquickowl]

@_abi_ @jxnlco @atbeme same

[Ben Cipollini @mrquickowl]

I'm going to be watching @OpenAI dev day live. Can't wait! Excited to see so much energy around it.

[Ben Cipollini @mrquickowl]

@Galileo_AI an app for using an ai chatbot to generate and visualize halloween costumes.

[Ben Cipollini @mrquickowl]

@Galileo_AI OMG I'm in. Making this app screen was as easy as typing "Language translation app".

Incredible.

[Ben Cipollini @mrquickowl]

The exciting moment has come! Galileo AI is opening up sign-ups for private beta.

If you want to start generating mobile app designs via chat for free, skip the line by replying to our Twitter thread below.

Be more like Alex and use Galileo AI for your next tight deadline.👇😃

[Stitch by Google @stitchbygoogle]

The moment is here! Galileo AI is opening up sign-ups for our private beta.

If you want immediate access and to start generating designs for free, reply to this thread to skip the line.

Be more like Alex and use Galileo AI to crush your deadlines. 👇

[Ben Cipollini @mrquickowl]

@Galileo_AI This is amazing! Love to test this - I know the pain of having limited design resources, and this gets the team started light years ahead.

[Ben Cipollini @mrquickowl]

I just watched the 3 hour Senate Hearing on AI.
The key concerns: Copyright, Influence, and Jobs.
The key suggestion: oversight from a new gov't agency with licensing and safety standards.
I like it.
https://t.co/7a4QpFiCYY

[Andrew Ng @AndrewYNg]

1/ Thrilled to announce: Our new course ChatGPT Prompt Engineering for Developers, created together with @OpenAI, is available now for free! Access it here: https://t.co/OaIpa6L2jn

[Ben Cipollini @mrquickowl]

@acgt01 definitely both :)

[Ben Cipollini @mrquickowl]

If you're looking to keep up on #AI developments, @theneurondaily has been an incredibly concise and deep daily read for me. Shocked they have 125 subscribers on Twitter (over 55K on substack!). Highly recommend. Note: this is my referral link https://t.co/l4pwgvuasD

[Ben Cipollini @mrquickowl]

We used to freeze brains to bring us back alive later. #ChatGPT and #LLMs make it feasible that we could recreate ourselves through our recorded outputs. Those who record and produce will live on - through others, personal growth, and now the possibility to reproduce their minds.

[Manolis Kellis @manoliskellis]

Such an honor to be back on @LexFridman to talk about #humanity, #AI, #intelligence, #ChatGPT, #DeepLearning, #DrugDiscovery, #Genomics and the #FutureOfWork.
Lex is an amazing human being, curious, driven, smart, humble, kind, and unapologetically open.
Thank you Lex for the many conversations, your thought-provoking questions, and your love of humanity. I always learn so much from you, your guests, and your insights 🤖🧠🫀🧬

[Ben Cipollini @mrquickowl]

@gdb #AutoGPT is a bit more worrisome; the level of human involvement is lower. It's definitely past time for regulation; wish we had folks with expertise on the government side to help guide it!

[Ben Cipollini @mrquickowl]

Great talk, including demo of #AutoGPT -like version of #ChatGPT. Even better Q&A by Head of TED Chris Anderson, driving hard on safety issues.
@gdb demo has humans defining goals and validating output. ChatGPT is a tool, but when will semantic goals emerge (or be programmed in)?

[TED Talks @TEDTalks]

ChatGPT stunned the world — but as @OpenAI cofounder Greg Brockman (@gdb) revealed live on stage at #TED2023, it’s only the beginning!

Watch the full talk + Q&A with @TEDChris on how ChatGPT came to be, plus Brockman's take on managing its risks: https://t.co/mcgdqtTiBg

[Ben Cipollini @mrquickowl]

@gdb It may already be there. #GPT4 system card shows emergent goals when system is empowered to act. These seem learned from the training text (e.g. the plots shown in the demo). Can turn if user is a bad actor (and this is being addressed), but main fear is different: a "rogue AI".