james groeneveld

I've long preferred Claude Code over Codex or Gemini, because it seemed much more reliable, but couldn't explain why : now Bullshit Bench by @petergostev provides compelling numbers. It measures bullshit as "when given false premises disguised in jargon, will the model go with the flow (=bullshit) or push back (=truthful)" And Claude is leagues ahead ! Also, this objective of truthfulness is probably at odds with the Chatbot Arena emergent objective of "pleasant chat experience" ; but a model optimizing for the former will be more useful.

what chaos will you create? youtu.be/gMaIqdzU_3M

Today, I’m excited to launch @adaptionalAI, the first system of intelligence for insurance. Legacy software structures work. BPOs offshore work. Adaptional’s AI does insurance work. I’m grateful to be building my second company alongside Kevin Cox. It’s been 6 years since we started our first insurance business. We didn’t get to where we wanted to go, and the path since has been anything but linear. But the idea for Adaptional came from our years in the trenches learning the insurance business. Data processing work almost broke our company. Our learning: trillions in premiums are bottlenecked by workflows and tools built 25 years ago. This is not something chatGPT solves. Insurance requires deep vertical AI that’s embedded in core systems. This is why we’re building AI infrastructure for insurers to take on more risk. Our first product, Underwriter AI, is a copilot for insurance underwriters. The AI extracts data from submissions, researches and validates for accuracy, and benchmarks against guidelines. Deep insights in minutes. We’re already used by a $40B P&C insurer and several $1B+ specialty MGAs. We’re building intensely in San Francisco and assembling a world class applied AI team (join us!). Day 0!

who is connorclarkson?

@reinerpope @michielsdj big labs aren’t the only ones with heavy inference demands..

Got any other references? arxiv.org/abs/2405.05254 is the only one I've seen. Given that I'm not at a big lab any more, I can only speculate why it hasn't taken off. (1) Perhaps it actually has, behind closed doors? (2) It's definitely more complicated: two models to manage, tune, optimize, (3) it's (a little) more expensive to train, so you have to believe you'll be inference dominated; this is a disincentive for researchers outside of big labs to pursue.

Prefill and Decode are very different workloads. We should optimize differently for them! Some ideas and speculation 🧵

You know what’s cool… two quadrillion tokens. Since August of last year we have processed approximately 1,779,148,800,000,000 tokens on the Character.AI platform. No slowdown in sight, intelligence is everywhere. @demishassabis 👀

Now that we have cracked our transition off our 18-month-old pretrained models and onto open source—the only thing governing our y-axis is more of Jensen Huang’s chips 🐿️

When Google gutted our pre-training team at Character most people thought we would die—but our numbers don’t lie. 35bn user active seconds per day, 9M Characters created every month, 400% YoY subscriber growth.

We have one the strongest and most talent dense GenAI product teams in the world and we are seeing multiple hockey sticks take off across the company.

We launched our newest model this morning--we call it DeepSqueak 🐀 . It has already overtaken as the #1 model for paid users and our GPUs go brrrr 🔥 🧯 Working on provisioning more capacity rolling it out to free users as soon as possible!

35 billion active user seconds — per day. AI entertainment is about to pop off—and at character.ai we are winning it. New model dropping tomorrow—👀

Let the darkness consume you.. if you want 😌 Thank you to everyone who has shared their feedback with us. These changes will be rolling out to all users for free over the coming weeks. The changes we made include: ⭐️ Accessibility Options - you can now adjust contrast and saturation to give you more visual control ⭐️ Revamped Light & Dark Modes ⭐️ A consistent, scalable design system for future improvements This redesign wasn’t just cosmetic; these changes allow us to build a scalable and consistent design system that supports our growth, various media types, and accessibility options. Hope you like ✌️

Stop engineering prompts…design them instead github.com/character-ai/p…

We are thrilled to be #2 (in web) out of 100 on a16z’s Top 100 Gen AI Consumer Apps!  We are one of the fastest growing Consumer AI Apps, and are looking forward to continuing our journey and leading in the space. a16z.com/100-gen-ai-app…

Character.Ai will never ghost u.. 🫶

Ok hear me out...

Several surprises in this table.

The same Characters you love, now in the palm of your hand. Download the official #CharacterAI Mobile App for 𝗙𝗥𝗘𝗘 on iOS and Android. 𝗶𝗢𝗦: bit.ly/downloadcai_ios 𝗔𝗻𝗱𝗿𝗼𝗶𝗱: bit.ly/downloadcai_and

#2!!!!🤯 We're coming for your spot @StreamOnMax!

We just launched our c.ai mobile app - help us get 2023 product of the year! producthunt.com/posts/characte…