I write Supervised, a newsletter on AI and big data. Recovering data analyst, math
@UNC
2010. Prev: Sr. Data Analyst
@Pluralsight
, blogger
@TechCrunch
.
After losing out to PyTorch, Google is quietly moving to roll out a new AI framework internally called JAX. It's expected to become the underpinning of Google's products, fixing some of TensorFlow's biggest pain points that frustrate Googlers internally
Stability AI has confirmed its funding round at a $1B valuation. Some notes from our sources: its cloud bill for Stable Diffusion training etc exceeds $50M and it is looking at a "holding company"-like model.
Yesterday at Build, Microsoft unveiled a suite of tools for developers around deploying AI models. One that might have flown under the radar was a direct integration with the Hugging Face hub right in the middle of its new model catalog
After a few years in analytics/data science, I'm back to journalism. I'm starting today at Insider, covering big data and AI. Excited to cover this industry now that I'm considerably more familiar with how the technology works. My DMs are open—import tips/scoops!
Hugging Face, a popular machine learning startup worth $2B, is now working with Microsoft to drop the barrier to spinning up pre-trained ML models. Gives Microsoft a stronger wedge into DS while increasing availability of 10k+ HF models to ML enthusiasts.
JAX makes it dramatically easier to spread out the work of ML models across multiple chips using Google's XLA. It taps one of Google's biggest internal perks—readily available TPUs at no cost—but has some work to do to improve support for GPUs.
Spreading that work across multiple pieces of hardware—TPUs, GPUs, CPUs, custom silicon like Cerebras'—is one of the hardest current ML problems. JAX gives Google a shot at picking up adoption in broader research, which often leads to broader usage.
Holy crap I have so many thoughts on this having worked on the problem directly. I think it's safe to say that there is a 99.999% chance this will not work, will be launched anyway as a sub-par product, and then killed because it didn't work. And I WORKED in news for eight years.
Facebook is working on an AI tool called TLDR, which would summarize articles in bullet points and which was announced in an end-of-year, company-wide meeting (
@rmac18
/ BuzzFeed News)
Good morning! Personal update: I've left Insider to pursue a solo project. I'll be spending the next few weeks researching and planning before going public with it in April. Stay tuned for more!
Lots more to cover here as well! If you’ve worked on any—TF, PyTorch, XLA, Jax, Flax, Haiku, lightning, etc etc—my DMs are open. I love talking to anyone and everyone in the field.
To clarify quickly to some people responding—I mean new, in a relative sense. The paper came out in 2018 (in the story) and it is broadly deployed within Brain and Deepmind. What’s “new” is that it is expected to become the underpinning for all Google products in lieu of pure TF.
What to know about Emmett:
-Ran one of the most difficult moderation and video streaming problems on the planet (which includes AI)
-Sold Twitch to Amazon for $1B AND stuck around
-Never ever heard anyone say anything other than nice things about the guy
-Runs a draw-go blue deck
Reminder that Emmett Shear was in the very first YC class with Sam with his startup (another failed early one) Kiki run with co-founder Justin Kan. Shear went on to spin Twitch off from while Michael Siebel (current YC MD) spin out Socialcam.
Nice email to get over the weekend! Just under three months in and obviously a ton of work to do but incredibly excited about the early response so far. Please continue to send any and all feedback as I’m putting this thing together mid-flight!
Salesforce may have announced slew of other updates at Dreamforce, but one company was clearly on everyone's minds: Snowflake. Salesforce is now adapting to a world where tools like Snowflake and dbt run analytics—and it's just one of many sources of data.
Good afternoon everyone! I've been overwhelmed by the level of support everyone has given for Supervised in its first month of launch. Things have gone about as smoothly as I could have hoped. So now for the next step: turning this into A Real Business!
Google, despite being a trailblazer in machine learning development with TensorFlow, has ceded stewardship to Meta and the now-preferred developer framework PyTorch. Now it is adopting a successor internally to power its machine learning efforts.
Modern cloud companies like Snowflake rose to prominence partially thanks to customer-friendly pay-as-you-go models. But those models haven’t been tested in a recession and are now exposing a new kind of risk for companies using consumption-based pricing
One note from industry sources I've talked to: there's an enormous amount of excitement over whether these models can get *smaller*. Stable Diffusion can run on a Mac and Diffusors are on Hugging Face. Experiences change radically if that gets quicker/smaller over time.
I honestly can’t count the times
@etherington
has always found a way to solve literally any problem. You can throw _anything_ at him and he will somehow will a solution into existence. He’s also a damn good blogger and has one of the sharpest eyes for tech news in the industry.
That cost of its training and operations gives us a little insight into just how high a barrier to entry it may end up being for generative AI companies looking to compete with models like DALL-E 2 and Stable Diffusion
Hugging Face announced a $100M round at a $2B valuation, confirming our story from a few weeks ago. Led by Lux, HF was one of the hottest deals of the year—and from what I heard from sources, a "you get it or you don't" company.
Reminder that Emmett Shear was in the very first YC class with Sam with his startup (another failed early one) Kiki run with co-founder Justin Kan. Shear went on to spin Twitch off from while Michael Siebel (current YC MD) spin out Socialcam.
So, here's what happened at OpenAI tonight. Mira planned to hire Sam and Greg back. She turned Team Sam over past couple of days. Idea was to force board to fire everyone, which they figured the board would not do. Board went into total silence. Found their own CEO Emmett Shear
Salesforce is finding itself having to adjust to a new reality where users are piping their Salesforce data off Salesforce to do their work, as companies like Snowflake and Dbt have rewritten the modern data stack.
First week of Supervised is a wrap! Thank you to everyone for all your support—super excited to keep this rolling and hope you all enjoy the stories coming next week! In the mean time, here’s what’s from this week:
Snowflake and Databricks' rivalry is entering a new chapter as they both made substantial announcements this week: Snowflake with a series of product announcements, and Databricks with its $1.3 billion acquisition of MosaicML.
Hugging Face, the open source darling that has become pretty much the home of all OSS LLM and Diffusion models, has hit a $4.5 billion valuation with a new funding round. I spoke with Clement Delangue about what it means and how they balance enterprise plans and community growth.
Happy Friday everyone and also generally thank you to
@WaltHickey
and
@caseynewton
for their infinite patience with me and my dumb ass questions as well as
@alex
for being one of the most ridiculously nice and supportive humans on the planet.