My book with
@eigenhector
is out! This is a practitioner focused book on all the technical aspects of building recsys.
This is a wide and deep coverage of the kinds of things Stitch Fix and YouTube use in their production recsys.
And yes it touches on llms
@xbaiIie
Started to laugh at my mom’s new apartment when I saw these decals and asked her about them, her answer: “well, I can’t afford any art really, and if I hang the pictures I have without frames they look bad, these were only a couple dollars.” 😔
I’m proud of Part 3 of our LLM series
I’ve been to over 100 LLM dinners in the last year, with everyone from Ai CEO at Salesforce, to ML lead of 100k person company struggling to figure out AI.
People need a strategy, so we wrote about how to build one:
Proud to bring you:
A Year Building With LLMs, a three part essay published with O'Reilly
We get into the weeds on what it takes develop incredible LLM powered applications
Advice from:
@eugeneyan
,
@charles_irl
,
@HamelHusain
,
@jxnlco
,
@sh_reya
and I.
Thrilled to announce that today I start my next adventure as the Head of Data Science at
@weights_biases
.
I’ve been a fan for years, and I’m thrilled to have the opportunity to work with the team and drive new capabilities.
Exciting things to come!
I'm excited to announce my new open-source project for LLM prompting!
They say the best products are things that you personally want, but nobody else has built yet.
So I present to you Yapping:
Ok data science/applied math Twitter:
Consider the Lp norm for p>2 but not the infinity norm.
Tell me p and an application space where this norm is right/good/useful/captures some important structure.
Looking for references too if you have em.
This was about 289miles over about 2.5 years. I would also like to add that my dog, Ravioli, was there for ~90% and my wife about ~65%.
I highly recommend this kind of project. If you’re already doing regular rides and hikes, it’s a worthwhile addition.
@technecat
@xbaiIie
I agree; her space is temporary during a rough transition but I was glad that she's making it more personal and livable despite. I think seeing her space for the first time was enlightening to me how different things are for her. but I digress...
Data science blogs be like: "Here's a beautiful cartoon to easily explain dot products".... lower on the page... "here's 20 lines of computation of Lagrange multipliers without a single picture or explanatory sentence".
Excited to bring you my recent news; I’ve joined
@_hex_tech
to bring Ai and ML capabilities to the platform
Initially, my team will be focused on our product called Magic — our NL-to-code capability
So if anybody asks what I’m up to, just tell ‘em I’m over here pondering my orb
Consider for a moment: you're a lowly ML engineer who just put your cool new model into prod.
How often should you retrain your model? What factors do you include in this analysis? Where do you start as a back of the envelope rule?
So you wanna log your rich data to Tables and have a quick UI integration to apply Dimension Reduction and Projection?
We gotchu:
Check out
@weights_biases
' new Dimension Projector! I'm super excited about this feature. :D
AI Data Visualization 📈
About a month ago, Grant Marvin and I set out to make AI generated charts generate faster while not sacrificing quality 📊
You could say it was a success.
Our new blog at
@_hex_tech
about Magical AI Charts in 1/10th the time 📉
The AI team at
@_hex_tech
just shipped two new features which make iterating on code with AI the best experience I've seen (and trust me, I've looked).
♾Continuous Edits: Rapidly iterate on changes in a single cell
🆎Diff Based Edits: AI to only changes the code relevant for
Are you–or someone you know–a bright and enthusiastic Data Scientist interested in the MLOps or ML Tools space? Wanna get nerdy about how ML Engineers train/deploy/experiment 📈 with models?
I'm hiring! Feel free to reach out via DM or apply directly!
Are you ready for round 2?
@eugeneyan
@charles_irl
@HamelHusain
@jxnlco
@sh_reya
and I are back with Operations
People loved Pt 1 of our Year of Building with LLMs series; for those of you who are ready for the next level of leverage, buckle up:
Some have asked so:
Next goals (on foot):
- Albany (1% to go)
- Emeryville (3% to go)
- El Cerrito (80% to go)
- Piedmont (99% to go ;-P )
- Oakland north of the highway (~75% to go)
- BA Ridge Trail (~65% to go but will likely redo segments)
Bike goal:
- Berkeley (35% to go)
Proud to announce we have attendees from top tier companies like:
- stealth
- unemployed
- N/A
Come learn and geek out with a bunch of awesome people, who are building with LLMs!
They wouldn’t let me name this
“So you think you can prompt”.
I’m hosting a mini conference. June 22. Come talk to other builders. Some share some hacks.
Saturday before
@aiDotEngineer
conference.
Today my math buddy at work taught me about Karatsuba Polynomial Multiplication which allows you to get a log time complexity for multiplying two degree N polys.
I really enjoy these types of clever tricks(although strongly believe nobody should ever be expected to know them)🧵
Mastodon really isn't that complex! When you sign up, you simply consider the unique X-scheme π: X_Z → X such that its corestriction to each open affine subscheme is isomorphic to the corresponding affine blow-up.... 🧵1/243
Mastodon really isn't that complex! When you sign up, you simply prepare a number of Spells in your Spellbook equal to your intelligence modifier plus your current level. You can then Cast them so long as you have Spell Slots equal to or higher than the level of the... 🧵1/243
I'm giving a paper-review talk this Friday @ 10am PST as part of the
@ml_collective
series.
The paper I'll be discussing is the Nature paper about Math + DL collaboration!
Representation Theory, Graph Neural Networks, AND Inference attribution!
Details:
@ErinChack
@jonnysun
My life version:
Lame date I
2 Lame 2 Date
Lame date: Conversational Drift
Lame & Dating
Lame Five
Date & Lame 6
Date 7
The Fate of the Lame
I am once again asking you to check out a blog post about Recommendation Systems. :D
I wrote about implementing Regularized Matrix Factorization from Scratch in JAX. Check it out!
Very excited to announce my newest project!
💯 The Route to the Top 100! 💯
The traveling salesman problem applied to finding a bike route that climbs the 💯 hardest road climbs in the continental US!
h/t to
@randal_olson
, pjamm,
@ridewithgps
, and
@Strava
@sigfpe
Do you offer package deals? I'm interested in all naturals such that after a finite number of applications, of the following rule, the value is one: if the natural is even, divide by two; if odd, triple and add one.
This seems like small set so I expect a good price.
A couple years ago,
@ericbunch0
& I found a paper discussing the combination of Topological Data Analysis, time-series, and CNNs; we thought it would be fun to read and implement, and maybe get inspired...
After a long journey here's our paper:
🧵 (1/7)
What's the version of "never work a day in your life" but for "work incessantly at all hours; wake up in a frenzy afraid to take too long to get to work; stay up too late trying to get a bit more reading done; because your always doubting competency and hoping you can catch up?"
You: It's cool that there's a trick for co-occurrences, but it's only for pairwise incidence
Me: Hold my tensor
My recent fun note on tensors to encode co-occurrences for higher order tuples, and why it's the same trick as for pairwise:
lil thread (1/6)
They wouldn’t let me name this
“So you think you can prompt”.
I’m hosting a mini conference. June 22. Come talk to other builders. Some share some hacks.
Saturday before
@aiDotEngineer
conference.
There will be an IRL extension of the LLM Conf hosted by
@BainCapVC
on June 22, 2024 in SF, led by
@BEBischof
.
There will be a "lightning hacks" session. Attendees are required to bring a hack. Bryan will reply to this tweet with what this means.
🆕 pod: with
@BEBischof
of
@_hex_tech
!
On putting AI Magic into Notebooks, and how RAG is actually a recommendation systems problem.
Also an
@AAAzzam
hot take: LLMOps is more like an "iron mine" than a "gold rush"!
@JulietteBruce12
During a colloquium by a famous mathematician who started with “consider a space”, I raised my hand and quickly asked “what’s a space”. He looked at me with disgust and said “something that *looks* like a space.”
@dimvar
I use
@wandrer_earth
! If you’re using a gps tracker already it hooks into that for the data and it handles the visualization and the aggregation.
Closing this chapter of my career after almost three years at Stitch Fix.
I’ve met some incredible friends and colleagues who’ve taught me so much.
More info on my next adventure soon.
TIL(and I'm a bit embarrassed) that probability density and probability mass functions are actually intended to have different meanings.
Mass for discrete distributions, density for continuous. 🥺
I got the opportunity to expound on MLOps with amazing coauthors
@HamelHusain
and
@dk21
.
Check out this no-BS overview of what's what in MLOps. It's a big topic – we tried to make clear why the field is so important, and what can go wrong.
.
@of_rude
at some point in each episode I think each of the characters is my favorite. I genuinely cannot even settle on a favorite. Most recent episode was freaking hilarious as usual.
Currently the best resource on applied LLMs… and it’s totally free.
This has talks on basically every facet of building LLM applications, in many cases the single best resources in those areas.
If you remember our Applied LLMs course, you'll love this. Today, we are making all these resources available for free to everyone! 📚
We did extra work to add learning tracks, resources, and notes to each lesson to maximize your learning. Link in next tweet
Today we are playing a DND session enhanced by AI services.
Gpt, Dall-e, and
@midjourney
for “assistance”.
This thread will have some highlights.
#gptrpg
@becky_k_warren
@Ayliean
@nhoskee
And here is my final zine that I started on during the workshop!
~~~~
Math & Tartans!
~~~~
A short booklet explaining how to make the repeatable pattern for a tartan, and an exercise to color one yourself.
For folding instructions:
#mathzine
#MZF2021
The team at Weights & Biases have finally cracked the barrier; we're rolling out an AI that can write Machine Learning blog posts for our community!
Read about our approach here:
Consider this an invitation to start thinking more deeply about your retrieval problems.
When Greg previewed this with me I was so excited that someone was writing publicly about this a bit more coherently
5 Levels Of Text Splitting: Semantic Splitting
Goal: Use embeddings to find chunks in our raw text
Experimental method using embedding distances between sequential sentences to tease out 'breakpoints'
If we find a good breakpoint, make a chunk
Check out the full video on YT
It's disappointing to see this topic as a punchline for some commentary on cancel culture.
Palestinian genocide is not an appropriate hat to hang on your straw man about some "purity of mathematical results" claims.
*first day of grad school in genetics*
Other student: “so, what do you want to work on for your research”
Me, stoically: “like Russian dolls, but avocados”
I have a few podcast shoutouts today because honestly I’m really impressed.
@fishnets88
is Killing it lately. Several guests in a row that have been fantastic. Check it out here:
@latentspacepod
had one of my favorite episodes ever of their’s with