Bill Chambers
@bllchmbrs
Followers
1,634
Following
851
Media
186
Statuses
1,858
👷 ✍️ 1x indie exit 1x O'Reilly author Prev: 🚀s ➡️ Anyscale, Databricks, $PCOR Talks about Startups, Data, AI
Joined November 2014
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Liam Payne
• 2205550 Tweets
Hobi
• 865055 Tweets
ブロック
• 375989 Tweets
Twitterくん
• 274532 Tweets
西田さん
• 207066 Tweets
西田敏行さん
• 191042 Tweets
俳優さん
• 148844 Tweets
ツイッター
• 103575 Tweets
Bluesky
• 91129 Tweets
東京の自宅
• 87781 Tweets
利用規約
• 70060 Tweets
ミュート
• 67264 Tweets
規約変更
• 64662 Tweets
ブルスカ
• 54346 Tweets
AI学習
• 54291 Tweets
イーロン
• 53660 Tweets
#ArrestMaulanaRashidi
• 53335 Tweets
महर्षि वाल्मीकि
• 51643 Tweets
pixiv
• 46120 Tweets
ぐりとぐら
• 43250 Tweets
中川李枝子さん
• 36796 Tweets
블루스카이
• 27622 Tweets
他のSNS
• 20582 Tweets
池中玄太80キロ
• 20074 Tweets
タイッツー
• 19542 Tweets
#INDvsNZ
• 19258 Tweets
ドクターX
• 17236 Tweets
世田谷区
• 15088 Tweets
Misskey
• 14662 Tweets
探偵ナイトスクープ
• 10381 Tweets
Pinned Tweet
Another important milestone on the solopreneur journey. One of my projects was acquired!
It was a modest acquisition so nothing to write home about, but it's validating to have built something from scratch, based on conviction, and getting that thing bought.
1
0
9
A gentle introduction to
#DSPy
...
I started kicking the tires on DSPy a couple of weeks ago.
I struggled. It was complicated and... I have grug brain.
This video is a GENTLE introduction to DSPy.
Now, don't get me wrong,
@lateinteraction
and
@stanfordnlp
are responsive on
8
37
306
The full "Gentle Introduction to DSPy" article is up:
A follow up to the video:
I've got a bunch more content coming here on this topic, I'm going to be breaking down all the key functionality of DSPy for others.
I'm just
4
39
227
DSPy, DSPy, DSPy!
All this week, I'm going. to be posting about my learnings about DSPy.
DSPy by
@lateinteraction
and the
@stanfordnlp
is an amazing tool in the AI ecosystem that not enough people know how to use.
Let's dive into WHY that's the case.
FIrstly, DSPy is a bit
6
24
205
DSPy suggestions are incredible. Think this needs a deeper tutorial - anyone interested?
13
3
153
Why do people get confused by DSPy, part 1:
This confused me when I first got started. Let's dive in.
5
10
128
I dove very deep, as deep as I am able on function calling in DSPy this weekend.
Here's all that I learned:
(all references for your own education / reference are in the comment below)
1. DSPy doesn't integrate with native function calling capabilities.
For instance, if you
14
14
125
who is working on agents in DSPy?
Anyone try to implement this sort of thing?
11
13
110
Something I've been experimenting with
#DSPy
how do we effectively extract the prompt to use in
@LangChainAI
or
@llama_index
?
I want to use DSPy everywhere, but not always feasible. Here'sa simple function to do just that .... explanations follow in the thread
8
10
95
Sketched out, in a single slide, some of the ways that DSPy can improve and write prompts.
This is all in the cheatsheet + docs, but wanted to summarize it (if only for myself) hope it helps others.
Fine tuning is missing (but more involved)
4
10
93
Alright, finally dropping this tutorial on DSPy Assertions and Suggestions.
To be honest, this is my new favorite feature of DSPy. It's so simple, but so powerful.
It's a little complicated to understand at first, but once you've got it - you'll see the opportunity.
Let's
4
14
75
After almost a year of work, Spark: The Definitive Guide is finally in production!
4
34
66
Why don't more people grok DSPy?
Because there's no company behind it (yet).
@ApacheSpark
suffered the same problem.
@matei_zaharia
,
@alighodsi
and founders TRIED to show companies the potential.
Eventually they had to start a company to prove it.
That will happen again.
10
6
71
DSPy Prompts
(DSPy week continues
@lateinteraction
and
@stanfordnlp
!)
Yesterday we took an end to end approach to building a small
#DSPy
app.
Today, we're going to DIVE DEEP into DSPy Prompts (ok, they're called signatures).
This video breaks down a couple of different ways
3
11
66
First things first. Let's talk about Prompts in DSPy.
Prompts in DSPy are known as Signatures.
A signature can be created in two ways (that I've found so far)
- Via a shorthand
- Via a class based definition
You can see both of those in the images below.
They "compile" to the
2
5
60
As I've spoken with so many people trying to understand DSPy, I've realized a big shortcoming.
People don't understand Evals. Should I do a deep dive on that next?
Evals Week?
10
4
59
So many shortcomings in
#DSPy
.
Not DSPy's fault, it's early yet.
It can be hard to use.
hard to learn.
It has gaps.
But when you get something down in like 15 LoC that you coudn't ever imagine doing before, you can't help but grin.
5
4
57
Prompt formatting in
#DSPy
here are three different ways to create prompts (ok signatures) in DSPy and control exactly what comes in and out.
First is the shorthand.
Second is class based.
Third is fully customized.
Use the third to work with multi-line inputs and outputs.
1
3
55
Great overview of DSPy and some of the motivations.
Reminds me SO MUCH of
@ApacheSpark
DAGs and SparkSQL Codegen.
0
5
49
DSPy (ad hoc) Meetup Notes from the Ad hoc Meetup on 5-23
attendees ~ 100+
cc
@lateinteraction
,
@matei_zaharia
,
@arnav_thebigman
I'm breaking this down into 3 (ok...4) sections:
What impressed me
What opportunities we have
Where we should go from here
misc thoughts
What
3
11
50
DSPy, DSPy, DSPy.
Why am I talking about
#DSPy
.
DSPy, or a system like it, will underpin ALL AI systems at some point.
It's early, it's sophisticated, but it will.
Once you understand it, you see the potential.
Prompt engineering is begging LMs.
DSPy is managing LMs.
1
6
49
More and more people reaching out about DSPy - even old friends :)
I wanted to write out, simply, my assumptions and why I took the time to experiment so much with it and share those results.
Assumptions:
1. Most apps are going to start with a big expensive model to ensure
3
3
45
#DSPy
week continues with answers to some great question from some followers (I read every comment) motivated this video...
What's the topic?
a Gentle (Conceptual) Introduction to DSPy (
@stanfordnlp
,
@lateinteraction
)
This is a departure from all
#DSPy
content so far and I
4
9
46
"my advice would just like build as many things as possible. Like, I think it's still really early in the space. No one really knows what they're doing to some extent. Like, it's a bit weird to say, but there's so many things to like discover. So I would just say like, build as
4
10
44
Caught this on AI news.
Couldn't believe my eyes.
#DSPy
cough,
#DSPy
...
DSPY gives you literally a lot of great examples... amazing to see others calling this out as so critical too.
DSPy+Haiku is such a great combo.
"In a lot of cases you may think you need fine-tuning, but you really don't – you just need to give the AI a lot of examples."
Nice to see some of the very practical advice I gave to the Computer Science Club at
@txst
validated by proper research just a few days later.
1
4
32
2
2
38
If you're using DSPy suggestions, use (or at least try) Haiku. The model is so responsive to feedback so far.
4
2
37
@TechFundies
Let's just summarize...
3) $Databricks is going to negatively impact $SNOW growth rates: lots of customers moving more $SNOW data / ETL workloads into $DB to save money, and $SNOW is way behind in AI.
Artic embeddings and model are interesting, but Snowflake can't even host
0
6
35
Great comment from someone about
#DSPy
Signatures...
SIgnatures aren't prompts.
They're not prompt templates.
They're their own thing.
Going to noodle on this and come up with a better way to explain them.
What do you think?
9
1
36
SO many great questions yesterday as I kicked of "
#DSPy
Week"!!!
- Why should I share about
#DSPy
?
- How does
#DSPY
compare to prompt engineering?
- What are the benefits of DSPy?
My upcoming video covers all that and more. It tells the story of an
#AIEngineer
named Nancy.
1
4
33
On Tuesday I finally received the email I've been waiting for. After 518 days, Spark: The Definitive Guide by myself and
@matei_zaharia
is finally going to print. Physical books will finally be available in just a few short weeks!
2
4
31
super impressed with the DSPy community - spoken with 4 people individually on things to change / improve - all from different companies and places.
A sign of a healthy community!!!
2
0
27
Most people don't know that I studied HISTORY in undergrad.
A long way from reading cutting edge AI systems.
If you're just getting started, stick with it.
"It" being whatever interests you.
Your life won't be linear.
Like all the code I'm tracing to understand DSPy Agents 🤣
3
0
24
Played around a bunch with function calling in DSPy last night.
Not teachable yet.
It's a lot different than something like instructor or
@llama_index
- at least it doesn't have the simple decorator style that I'm used to....
Should have some info to post tomorrow!
1
0
23
Read my blog post on getting a gentle introduction to
@ApacheSpark
Want an easier way to write an
#ApacheSpark
application? Check out this introductory tutorial:
0
18
20
0
1
20
Why do people get confused by
#DSPy
, part 2:
(Part of DSPy week!)
This is the biggest gap ...
Why do people get confused by DSPy, part 1:
This confused me when I first got started. Let's dive in.
5
10
128
1
2
23
Interesting to see more information coming out about fine tuning (link in the comments)
DSPy, DSPy, DSPy - if you're using examples / in context learning, you might be on the right track
Not sure any of this is new info, but interesting to see the research!
2
4
22
so I'm writing some DSPy and I'm thinking to myself... man it'd be great to hint to the program how to improve - just to give it a push in the right direction.
(me reading the docs): holy shit. that exists.
but then, it doesn't work 😭
3
3
21
dove really deep into suggestions and assertions in DSPy this weekend - anyone use these / familiar with them?
Will throw up some tips and tricks if it's useful for people!
3
1
21
Fuck it, let's make it happen.
I want to talk to others about DSPy.
This community is fantastic.
I want to setup a ... DSPy live chat on Thursday at 11:30AM PT
Free for anyone and everyone.
We talk everything DSPy.
What works.
What don't.
Anyone in?
Yes
92
No
4
10
6
22
I know i say this like every week but my new favorite feature in DSPy is suggestions.
the possibilities are endless.
For instance..
You could create model that is optimized to provide optimal suggestions to fix problems in your main model. And doing this, wouldn't be hard.
3
2
20
DSPy is an amazing tool because it gives engineers the power to move from model provider to model provider in. a PRINCIPLED WAY.
Engineers just don't know it yet. I want to change that.
2
1
19
I started experimenting with a new application called
@getclearspace
and I adopted and paid for it immediately - here's why
@getclearspace
works by INCREASING friction with apps
Adding friction is the best opportunity for us to regain control from apps designed to sap attention
1
4
17
Some relevant links for those that want to dive deeper.
Code:
Learn by Building AI Newsletter:
Learn By Building AI Youtube:
0
1
18
A gentle introduction to
#DSPy
...
I started kicking the tires on DSPy a couple of weeks ago.
I struggled. It was complicated and... I have grug brain.
This video is a GENTLE introduction to DSPy.
Now, don't get me wrong,
@lateinteraction
and
@stanfordnlp
are responsive on
8
37
306
0
3
17
alright work with me here.
You want to build a model per user / per customer / per whatever.
You can (1) train, (2) fine tune, (3) DSPy tune few-shot to optimize for that user.
Is there any real alternative besides (3)?
(1) and (2) seem prohibitively expensive.
what are the
8
0
16
The absolute biggest gap in DSPy in my mind is working with multi-line inputs and outputs. I cannot get it to work well.
I'm trying to generate large bodies of text and it's just a constant battle.
Has anyone used DSPy with multiple, large, example inputs and outputs?
6
0
16
I've been talking a lot of
#DSPy
but something that continues to fascinate me is fine-tuning. There's so much potential.
There's a great course coming up on the topic if you're interested by
@HamelHusain
and
@dan_s_becker
No affiliation, just looks great
0
0
16
First Learn by Building AI Email went out to the Subscribers!
Since it's DSPy week, All info from DSPy week consolidated for easy reference...
Like...
- A Gentle (technical) Introduction to DSPy
- A Gentle (conceptual) Introduction DSPy
- When not to use DSPy
1
0
16
This post is going to introduce a role the future. The Citizen AI Engineer.
Walking into the future, hand in hand with a robot.📷
This role is fundamentally nontechnical but that heavily uses AI to solve business problems.
1000s of new products will be built to service this
3
6
14
kinda feel like dspy is the gateway drug to fine tuning
or is fine tuning a gateway drug to dspy?
2
2
14
me n the crew after we finally got our DSPy programs working with no bugs
0
0
14
I feel like when working with DSPy ... sure the docs need some help but most of my problems are actually data parsing / quality issues.
how are people debugging this sort of thing?
4
0
14
Here's something people are missing.
When you're hand tuning a prompt, you're EFFECTIVELY TUNING THE WEIGHTs.
If you define a "program" or "optimizer" to tune those weights for you - magic happens
-
@lateinteraction
on a webinar
@llama_index
6 months ago.
2
2
12
@TechFundies
Baises:
1. Long $PCOR, worked there 10 years ago as a baby startup
2. Neck deep in Data + AI
One thing that I think is under-estimated (although I don't understand the extent of investment) is the data/AI growth potential.
Procore has so much data locked up in their systems.
1
0
9
I really thought this Chat-GPT4V model was overhyped. After using it this morning, I'm not so sure...
Is this the era of personalized software?
I took a mockup I built and asked it to style it and code it up for me. It did it. On top of that, it helped me build a full featured
3
2
11
That system performs OPTIMIZATION.
In order to optimize, you need to have a metric.
In order to have a metric, you must have. away of defining "good".
In order to define good, you've got to understand your problem.
No shortcuts.
2
3
13
Firstly, DSPy is a SYSTEM. It's not a prompt hack. You've got to think thoroughly about your problem - no shortcuts.
This is certainly something people refer to when they talk about the "learning curve"
@lateinteraction
This is so dense. Please share more elementary examples too. I want to try DSPy but my small brain cannot comprehend what's going on under the hood b/c of abstraction that I'm scared to try it for my business for fear of bike shedding.
2
0
3
1
3
13
People COMPLETELY MISS how important documentation is to a brand.
The first comment ...now anyone reading this thread on reddit is going to have a poor product perception.
Documentation.
Is how developers.
Experience your product.
👏Documentation👏Matters.👏
3
2
12
@theandreboso
This makes me think about a concept I'm going to implement for my next project. Something like "The Reverse Product" basically you do everything BUT build the product.
Marketing.
Funnel.
Offer.
Pricing.
Website.
Landing Pages.
You find that paid channel.
But avoid building
2
1
12
what are people actually using for web search to feed LLMs? Any recommended providers?
5
2
11
Come see me speak at
@SparkAISummit
in London! Going to cover the best practices we see
@databricks
for successful data science projects.
1
4
11
DSPy ad hoc community meetup -
I didn't record this one - but it went super well - so many amazing ideas and an awesome and friendly community...
will explore recording the next one but will send out notes to the
#DSPy
community here shortly
User
1
0
11
@WolframRvnwlf
For those that want to read:
Enjoy!!!
A gentle introduction to
#DSPy
...
I started kicking the tires on DSPy a couple of weeks ago.
I struggled. It was complicated and... I have grug brain.
This video is a GENTLE introduction to DSPy.
Now, don't get me wrong,
@lateinteraction
and
@stanfordnlp
are responsive on
8
37
306
0
1
11
@lateinteraction
@stanfordnlp
In ~one hour, I'm going to share a ~30 minute tutorial that is a GENTLE introduction to DSPy. We'll walk through a simple example to help people understand the basics of DSPy.
1
0
10
The hard part about DSPy is getting your data right.
Much of this is just data engineering - more and more, the prompt issues, the failures, the edge cases, come from issues in the data...
Spark and data processing engines are your best friend...
2
0
10
alright, I'm finally writing that tutorial on DSPy suggestions I'll be sending it out on the LBB newsletter this week, sign up in the tweet below if you're interested in learning more
1
0
10
@hwchase17
I took the time to dive in a bit more deeply into langchain and update the chat your data repo:
1
2
10
@remilouf
@aiDotEngineer
This was a top talk
@aiDotEngineer
- super clear value prop.
TL;DR
1. Generate structured output, it's better.
2.
@OutlinesOSS
is the best way to do that with open models.
1
6
10
The second thing you need to do is define. a metric that objectively measures whether an output is "good or bad". This metric can be a simple boolean or even a DSPy program itself.
START SIMPLE HERE.
0
1
10
#Langchain
(
@LangChainAI
) just dropped a bunch of videos on their youtube about langsmith, tracking, and evals.
Lots of great content - from DSPy to an end to end Eval pipeline. Hope there's more to come!
0
2
9
Awesome to see that Spark: The Definitive Guide is
#1
New Book in Java Related Programming on Amazon!
Lots of people asking for copies! They're coming soon!
1
0
9
Amazing to see what the
@haizelabs
team is doing around red-teaming LLMs.
This discussion isn't top of mind as people are getting to GTM but the advanced teams are already obsessing over this as a critical part of their product.
If you're building with
Today is a bad, bad day to be a language model.
Today, we announce the Haize Labs manifesto.
@haizelabs
haizes (automatically red-teams) AI systems to preemptively discover and eliminate any failure mode
We showcase below one particular application of haizing: jailbreaking the
94
169
1K
3
2
9
So, what does that mean in practice?
The first thing you need to do to use DSPy effectively is to get some data with labels. Those labels can be almost any kind of label, but you need labels.
3
1
9
References:
- Function calling in openai:
- Function calling in mistral:
- Reference implementation:
- Discussion of Agents:
- Agents.
- reAct docs:
-
0
2
9
wow this discussion went way longer and deeper than I thought ...
3
0
8
is it 2017?
Did you get your hands on Snowflake Notebooks yet? Experiment with your data, run end to end data engineering pipelines, train ML models, and build cool
@streamlit
apps all in one place. No more hassle of creating and managing python environments, dependencies, and more!
1
5
32
0
0
8
Awesome to finally be able to share Structured Streaming running on
@Azure
@databricks
! Look forward to watching our customers continue to build amazing things on
@Azure
@databricks
!!!
0
4
8
@meta_learning
is my new favorite learning resource:
http://t.co/calzN5CsTm
It really cuts through the fluff
1
5
7
I've been using Cursor .so to write code for me using GPT-4 for a couple of weeks..
I am not a software engineer.
Here are the places I've found Cursor .so helpful and unhelpful:
4
1
7
AI LIBRARY REVIEW: Instructor
A less talked about yet potent application for LLMs is structured data extraction.
Structured Output > defaulting to "chatbot" applications.
However, not all tools are created equal, so let's explore instructor by
@jxnlco
🧵
What is Instructor?
1
2
7
Hey friends, ever since I left my previous company - I've been in build mode.
I've had 3 goals:
1. Discover what's happening in AI & ML at a deeper level
2. Build indie sized digital businesses
3. Learn and grow as much as I can
2/3 were successful - guess which ones. :)
1
1
7
The LLM market resembles the Big Data Space of 2015. It's all over the place, lots of vendors are jumping into the market, and high quality material is hard to come by.
1
0
7
gotta love when in user calls you get this from the customer...
"This is a really important problem to solve"
0
0
6
At
#StrataData
? Come to "Streaming Big Data in the Cloud: What to Consider and Why presented by myself and
@michaelarmbrust
!
11:50am–12:30pm Wednesday, March 7, 2018 in Room 230A
0
3
6
@ahmed_moubtahij
So I host an llmproxy server on my local machine. That makes it so that I can use any (almost) in DSPy with just the OpenAI configuration (and by changing the API base).
Makes switching to haiku or opus just a one line change - pretty sure it works with HF as well.
2
0
4
@carlvellotti
Anyone want me to AI this thing? I could build infinite interview questions with evaluations of the answers. Let me know if this interests you with a ⭐️...
2
0
6
@mckaywrigley
this. is. insane.
Serious question. what is copyright anymore? what's a derivative work?
If you can go to a dribble screen, input the screenshot, and get the site. Is that copyright infringement?
If you can take a screenshot of a website and tell it to recreate it... is that
2
1
6