Sam Whitmore
@sjwhitmore
Followers
12,558
Following
1,651
Media
155
Statuses
1,800
building @newcomputer . not a cat (or a man) in real life. I like to run a lot! @kensho @harvard @StuyNY
San Francisco, CA
Joined February 2011
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
#GrandConcertENGLOTinMiami
• 459604 Tweets
Usyk
• 453884 Tweets
América
• 431430 Tweets
オークス
• 132716 Tweets
Luka
• 96261 Tweets
Dallas
• 90479 Tweets
Mavs
• 83038 Tweets
Thunder
• 77168 Tweets
Chivas
• 63719 Tweets
Kyrie
• 57536 Tweets
上川外相
• 56868 Tweets
ゴーオンジャー
• 40449 Tweets
Shai
• 39824 Tweets
切り取り
• 29281 Tweets
Lively
• 25986 Tweets
ステレンボッシュ
• 20272 Tweets
BRILLIANTSUMMER WithKYLEDREA
• 18097 Tweets
Chet
• 15439 Tweets
メイドミドリ
• 13248 Tweets
#cgUT_福岡昼
• 12607 Tweets
ライトバック
• 11615 Tweets
Gago
• 11118 Tweets
言葉狩り
• 10545 Tweets
Chicharito
• 10117 Tweets
ChatGPT is not new technology, just a new interface. You can replicate it pretty closely from existing APIs.
All you need is:
1. a good prompt
2. a memory window
Here's how to do it using
@LangChainAI
: 🧵
33
181
1K
I’ve been building stuff w agents for ~3 months. Some scattered thoughts on the current agent hype:
On agent reliability:
before you get too scared or hyped about autonomous agents, try running one yourself
they have flashes of clarity and then tend to rabbit hole / get stuck
57
154
1K
when did the ML -> AI acronym transition happen?
used to be all "ML" back in the day
163
34
995
How do you make an AI chatbot smarter and more engaging? By giving it long-term memory of the entities it discusses with the user!
That's what
@devennavani
& I built for the
@scale_AI
hackathon (live today in
@LangChainAI
). Honored to be one of the 8 finalists!
How it works:
38
62
803
holy smokes i googled for this for a solid 20 minutes with no luck before asking chatgpt and got a perfect response WOW this was my first life changing moment with chatgpt
17
51
769
here’s what happens if you accidentally initialize an agent without any tools and give it a task
“Samantha, can you see this? I have tried to text you, call you, email you, and speak to you.” Chills
cc
@LangChainAI
45
69
765
We’ve spent months iterating on long term memory for LLMs and today we have something special to share!
Announcing our first product, Dot — an intelligent guide with incredible memory.
Read the full story here:
209
509
5K
49
26
679
had my 10th college reunion this past weekend & it reinforced something i've believed in for a while, but i think people in their 20s especially don't often realize
don't do things to impress other people. no one cares what you do. really, mostly no one thinks about you at all.
13
32
459
been telling everyone i was in "goblin mode" because i thought it was a good approximation of "living in a cave and coding all day" but it turns out it has a totally different meaning which i never googled. to everyone i told i was in goblin mode, that's what i meant
18
16
452
hot take: I don't think generative AI is a category.
it's new tech for building custom data transformations.
with minimal instruction, you repeatedly transform (almost) any type of input data into any type of output data.
this is a completely novel building block for engineers
22
23
339
Security: “don’t forget to empty out your pockets”
My pockets:
10
15
304
when i was a young engineer i was scolded for thinking "tech-first"
feels like AI hype has gone a little too far in that direction recently. agents / LLMs should be implementation details in a good product, rather than the headline
dont forget what problem you want to solve!
10
21
280
latency for LLM-powered apps is often ignored in public demo discourse right now.
even 1 second total feels too slow in practice -- your user has already looked away from the screen.
this completely changes agent design IMO
42
15
268
I'm so SO excited to be joining forces with
@jasonyuandesign
!!!
Jason has always thought years ahead of his time. I can't think of anyone more well suited to bring something really special and new into the world.
if you're interested in working together, please get in touch :)
So what’s next? Well…
I'm excited to announce that I'll be teaming up with
@sjwhitmore
to build something new. I can’t give too much away yet, but I will say this: It is powered by AI. It is the future. And it feels magical.
24
5
316
17
4
266
wow big breakthrough last night...just from banging our heads against the same concept day in & day out for more than a year and trying 1500 permutations of the same thing i think we found something.... really incredible?!
8
2
251
to all the folks at OpenAI - you built something so special. I am so sorry you’re going through this. What an awful 3 days. 💔
4
20
243
here's my talk from the Memory Hackathon today - i spoke a bit about some of ways we think about developing memory for Dot. always down to chat if this is something you're thinking lots about also!
15
19
216
we'll be sharing more about what we've been working on in seven days!
15
7
213
Last New Years, I made a resolution to run at least 4 miles outside every day in 2022.
365 days and 2468 miles later, I did it!
Things I learned:
1) Keeping a daily promise to yourself is empowering
2) Commitment compounds over time
3) Getting outside every day feels great!
13
4
209
here is a list of LLM ops / infra products I want to exist.
1) synthetic email / calendar data for a person with a particular user story profile that sits behind a mock google api -- e.g. "the email inbox & calendar of a busy tech VC"
10
12
202
open source audio gen model by -- and huggingface space here --
these guys know what they're doing! excited to follow along!
8
25
197
for building with LLMs i think the classic engineer advice of "don't reinvent the wheel" is backwards now.
the paradigms are so new that most of the time, you should think by yourself from first principles how things should be done before exposing yourself to existing patterns
10
18
189
this is what we are looking for in hiring at
@newcomputer
- if this resonates with you always feel free to reach out to me or
@jasonyuandesign
taste to us means a good sense of what is mid & non mid in all things as well as a healthy anger at midness & desire to fix it
taste > skills
taste seems more scarce these days, and increasingly differentiating in the age of AI where so much of skills-based productivity is offloaded to compute.
makes me think about the development of taste, and how we nurture taste for the next gen of humans.
95
196
1K
7
3
191
perhaps spicy but
there is incredible alpha in being a woman building a personal AI product right now
women are half the population. do you know what they’ll think of your product?
19
2
185
I’ve run 785 days in a row & always told myself I would stop if I didn’t think it was the healthy choice. Finally felt today like it was a bad idea & made the call not to run.
it’s been a wild journey…excited to see what I feel like after my first rest day in 2+ years 😂
24
3
182
people often misunderstand why chatgpt is so popular. it's not cause it is conversational or "chat-based" in its response.
it's because it answers your question immediately w/ all of the information you need in one place
5
3
172
we have built an entire system to birth synthetic users w/ a consistent backstory & have them use our product repeatedly over time in a way consistent w/ their synthetic backstory.
its so wild watching the AIs talk to each other
now bringing in more AI to autoscore it all
9
5
171
hot take: i don't want to talk to most businesses via a chat assistant.
instead i just want them to magically give me what i need when i need it and i want to look at my phone & computer less.
chatting for everything makes me look at my phone & computer more.
14
10
166
day 700 of run streak today... 4197 miles run (avg 6mi / day)
my life was in such a different place when i started this... it's been fun to have daily running as a constant. excited to see where it takes me
13
1
164
i love that openai just updated their API to reduce laziness & my reaction to that is completely normalized excitement
2 years ago if you told me I would be tracking the "laziness" of a given API it would have sounded bonkers
sometimes you have to stop & smell the surrealism
5
9
154
have you ever spent 20 minutes confused about spaces in bash scripts? i just have
TLDR
bad:
if ["$s1" == "$s2"]
good:
if [ "$s1" == "$s2" ]
11
6
154
i haven't tweeted anything in a while b/c we've been on a roll building.
it feels like we are closer to getting stuff actually working so i wake up every day and its all i want to do.
resurfacing to say we still are looking for someone to help on backend tho. work
@new
.computer
4
11
153
so much happened this year but starting off 2024 with the same 3 goals as 2023
be a good wife, build
@newcomputer
, & run at least 4 miles every day. the infinite game continues.
grateful to be living when I am with the opportunities I have
6
0
150
First look at Daemon (), a chat-based AI personal assistant with long-term memory about you. The following is a fresh conversation, no prior context.
12
14
146
excited to say that we've been able to build on top of gpt-4 for the past few months (one of the reasons i haven't shared many demos / screenshots). the model is *really* smart. thx
@OpenAI
!
twitter's gonna be crazy today but will share more soon :) --
here is GPT-4, our most capable and aligned model yet. it is available today in our API (with a waitlist) and in ChatGPT+.
it is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it.
1K
4K
21K
7
7
148
where is the LLM-powered app that simplifies GCP / AWS iam roles / permissions
give it to me now
i will pay $$
21
11
144
this API is honestly incredible. its like using serpapi + browserless / selenium in one step, but w/ the added bonus of more interesting ways to search (neural in addition to keyword).
I almost feel a bit torn promoting it via twitter b/c its such an advantage to use them lol
Soon, AI will search more than humans. Metaphor is the first search engine built for your LLM!
With a few lines of code you can do a Metaphor (or keyword) search and get back clean, parsed HTML for any content. Get 1000 requests/month free at (1/3)
29
56
438
2
7
140
hot take: i think the labeling of different techniques in AI engineering discourages fluid experimentation and makes it seem higher barrier to entry
don't worry about whether you're using CRAG or HyDE or whatever, then you get confused about the implementation and stop thinking
14
11
136
when u run away from dev day to go integrate everything as soon as humanly possible
dream day for people building ai-powered products
thank you
@OpenAI
1
3
131
started using yesterday & i'm blown away. it's how Copilot *should* feel. i'm completely off VSCode now.
great work
@amanrsanger
and
@mntruell
. two concrete examples:
1) it generated a new viewset for me using the correct schema of my django models that
7
7
124
it is time for the annual recitation of the digits
one day I’ll forget the first 100 digits of pi, but for now I know them, so here they are
happy pi day!
18
5
124
@hackathongoofer
@LangChainAI
btw to all u "maybe the women dont want to be there" haters -- the second most prolific contributor to langchain is a woman (me). UGH the world is tiring sometimes
18
2
124
I hit 400 days on my run streak today!
hit me up if you ever want to join in SF — I don’t go particularly fast most of the time, I’m a consistency gal
Last New Years, I made a resolution to run at least 4 miles outside every day in 2022.
365 days and 2468 miles later, I did it!
Things I learned:
1) Keeping a daily promise to yourself is empowering
2) Commitment compounds over time
3) Getting outside every day feels great!
13
4
209
8
1
122
i see a lot of travel booking examples for AI agents but keep in mind the problem of travel booking for a consumer is not an execution problem but a research / optimization problem
- if i go one week earlier are prices cheaper? what if i fly into this other airport?
- what if i
10
8
112
i subscribed to the blog of an anonymous engineer who is spending a season in antarctica and am so happy every time i receive an update in my email. portal to another world
5
9
113
the proximity of incredible trail running to SF is just so nuts
ran for 2 hrs, saw 2 other humans
3
0
111
who is building the runtime for openai's Assistants API?
was pairing with someone on a sample project that called one function & had calendar integration & ran into immediate issues deploying it -
1 - have to deploy something that handles oauth for calendar (ugh)
2 - have to
37
10
105
@nickcammarata
there is a tech female version of this - fuel from being ignored or prejudged for formative early parts of your schooling & career by your peers
equally strong if not more so
5
4
106
someone should build a dating app that embeds your answers to all of the personality questions & calculates top cosine similarity across the pool of people on the entire app to find your semantic match
25
2
104
i say this coming off the reunion because people were so genuinely happy to see each other as people and our jobs / accomplishments faded into the background as we caught up, which was lovely. to focus on laundry list of accomplishments felt small & petty.
1
3
98
this week i learned that some people hear an internal narrator in their head?! like a voice that describes things in full language?! this was news to me
i think in a compiled language that isn't full sentences, but some combination of visuals + words. curious if this is common
37
4
99
It's 2028. The Internet is ruled by AIs due to the overwhelming influx of generated content & the ease of inter-AI communication for completing transactions.
Every human has their own AI, and your AI goes out to cyberspace to run errands for you & keeps you informed of
12
2
94
I hit 500 days in a row of running today!
This started when I made a New Years resolution in 2022 to run at least 4 miles every day, & I've just kept it going into 2023
In the past 500 days I have run 3164.6 miles, avging 6.3 mi / day. I have not missed a day
9
1
94
my husband got me a Pokémon Blue emulator for Christmas cause I didn’t play growing up. I finally understand the appeal, it’s so fun!!
Started with a Charmander and caught a Pikachu (!) in the Viridian Forest, about to go into the caves
13
0
93
this is the most freeing feeling once you realize it, and i hope you realize it young. it frees you up to concentrate on goals & pursuits you truly find interesting... and paradoxically allows you to do the most creative & unexpected things of all.
3
8
91
stoked to be speaking at the
@aiDotEngineer
conference with
@jasonyuandesign
!
we'll be talking about new interface & behavior paradigms for AI-powered applications.
there are lots of other amazing speakers too. remote tickets are free -- or come see us live!
Our full schedule is now live!
Which are you most looking forward to?
2
6
56
2
8
91
the AI version of me went to this simulated party & was one of the 10 least popular people there lol - other AIs found me "too intense" & "overly focused on work"
awesome job
@edgarhnd
this experience was inventive & fascinating even if i now need to work on my social skills
i'm throwing the first ever AI simulated party.
it's 3 days long.
day 1 and day 2 are in the simulation.
day 3 you pull up irl to Mission Control in sf.
here's how it works:
1. every guest gets an AI character.
2. you customize it to your personality.
3. your character is
163
35
372
7
3
89
early yahoo search was a list of lists of sources by topic
google revolutionized all that with pagerank
now i just want a list of lists of expertly curated sources again
9
2
89
need a word for the feeling when you've been thinking about the same problem for so long & all of a sudden the answer becomes totally clear and so obvious you wonder why you never got it before, and you simultaneously begin to worry its obvious to everyone else
16
4
89
Spending 10 days in Saskatoon, Saskatchewan and the temperature has been ranging from -35F to -20F.
I’ve been running outside every day and have learned that for cold like this you have to be super cautious about not exposing any skin. Every day my eyelashes freeze
11
0
88
@kempton_timothy
sigh, taking the engagement bait...
its different for everyone but being married allows me to take more risks & focus more as compared to what my own alternate-universe single self would be able to do.
dont listen to one-size-fits-all twitter pundit romantic advice, people
3
0
87
grateful that I saw
@sualehasif996
's tweet about
@turbopuffer
.
our query pattern is many many small vector indices that are used sporadically. pinecone always felt silly for this - why should i pay to keep thousands of indices running when not in use? or should i just have one
5
4
86
Wow, this is how I imagine tweaking a generated image - feels so futuristic. I have always had so little patience for prompting image gen only with words
real-time is here.
with it, a new generation of AI creative tools is coming.
282
753
4K
1
3
84
I was inspired by
@317070
's excellent post on creating a virtual machine inside ChatGPT () and wanted to see if I could replicate the same behavior using text-davinci-003. (spoiler alert: I did!)
1
5
82
we are so excited to partner with
@LangChainAI
to organize a hackathon focused on long-term memory!!! it's on 4/6 in SF - hope to see you there!
will be giving a talk and sharing some of the techniques we've used in Dot :)
🧠Memory Hackathon with LangChain, , MongoDB and Anthropic
✨Sign up here: ✨
Long-term memory for LLM applications is one of the most interesting and important - yet also under explored - areas out there
We are THRILLED to be
2
23
141
5
8
83
this has lived rent free in my brain since i saw it last week, & i've told like 5 people about it organically. amazing creative vision
1/ Today we’re launching in Beta. Anyone can now sign up to create and edit rooms. Here's a 90 sec trailer, and some reflections...
95
161
2K
5
3
81
Super excited to share that I'm joining the
@southpkcommons
Founder Fellowship!
I'll be working on a protocol that makes it easy to use machine learning in your smart contracts. DMs open if this intrigues you! :)
14
10
81
I gave
@CurioBeta
to a 3 year old and it’s like actual magic
We read a picture book about the solar system and Grok would chime in with asides about each planet
And honestly just the fact that it says his name & ours & can chat back is wild - his eyes lit up
So cool
6
4
80
spent this morning adding in chain-of-thought examples to a bunch of prompts I was working on-- just tested the same tasks with & without chain of thought on text-davinci-003 and they're actually better without
7
5
73
i predict you will see everyone who built "chat for x" apps over the next few months realize this & scale back on conversational responses in favor of dynamically generated UI which fits the user's specific query best
10
1
77
Happy Pi Day! I memorized a bunch of digits in high school and still remember 100, so here they are. I have to recite them every year because what else am I going to do with this random knowledge?!?
7
0
74
reminding myself once again that it just takes self belief, persistence, honest self critique, & the patience & willingness to try something over and over again slightly differently (yet quickly!) & learn over time
3
2
71
This bridge from Cambridge to Boston is one of three places in the world where you can have a plane flying over a car driving over a train rolling over a boat
3
5
72
i now find myself programming by talking to Dot about how it understands how its supposed to do something and then telling it if its right or wrong / helping it see the task differently
AI for therapy? no. i am the therapist for the AI
4
1
72
.
@lachygroom
is the BEST! I’m lucky to have him as an investor
some of the most gratifying work I've done lately is working closely with
@sjwhitmore
on her new company.
if you're an engineering founder, or an engineer considering founding a company and want to talk through product (or really anything!) just email me! lachygroom
@gmail
.com
5
22
149
5
0
70
It would be pretty fun to make a Museum of the Internet where kids could log onto machines with software & websites pinned in time at specific years, like 1998 or 2002.
We need to preserve the memory of modems, Yahoo / geocities, AIM buddy profiles, early blogger sites, etc…
7
0
70
🧠 This week I worked with
@hwchase17
on adding ConversationChain to
@LangChainAI
.
A conversational agent needs to remember what you've said to it so far and be able to draw on that for further responses.
We added an optional Memory module to the Chain class to achieve that.🧠
🦜🔗LangChain version 0.0.21
🧠 Added in Memory abstraction (
@sjwhitmore
)
💬 Conversational chain (
@sjwhitmore
)
➕ Method for adding examples to ExampleSelector class (
@sjwhitmore
)
🐛 Bug fix for search class
🍂 Support for all params in the OpenAI wrapper
2
5
27
5
6
69
Agree with this.
Trying to make the LLM itself infallible will be a losing battle. even if you push the probability of wrong answer down to 0.00001% you cannot trust the veracity as a user.
Instead, I view LLMs as incredible tools for transforming / processing information.
I have claimed that Auto-Regressive LLMs are exponentially diverging diffusion processes.
Here is the argument:
Let e be the probability that any generated token exits the tree of "correct" answers.
Then the probability that an answer of length n is correct is (1-e)^n
1/
219
539
3K
7
7
68
By this, I don't mean to say ChatGPT isn't impressive -- a change in interface itself can be mind-blowing. I use it myself for search all the time now. I just mean to say that if you're a developer, you don't need for the API to drop -- you can start building today.
2
5
64
when you’re at dinner in SF and
@DivGarg9
is demoing and you have to leave so the demo becomes an AI booking your Uber. Thanks
@MultiON_AI
for this Uber ride!
3
4
67
I was lucky to be there in person for this. Kevin isn't exaggerating, it was truly a one of a kind demo. at one point he had some A/V issues (audio stopped coming out of computer) and it took like 4 min to fix. normally people would be fidgeting and lose interest, but people were
It’s time to unstealth the worst kept secret on a Twitter - that I’m building a company to give AI a soul -
@OpenSoulsPBC
Last week
@betaworks
, I illustrated just how we’re going to imbue AI with the indescribable essence of humanity.
Many people have called me crazy over the
79
124
731
2
7
67
2) a platform that allows you to tag & save good user interactions or flawed ones, then rerun them after a prompt has changed as an integration test
4
2
66
@alfongj
@LangChainAI
I kicked this off and switched windows and literally felt a wave of guilt when I looked back at terminal and saw it was doing this
3
1
62
A couple improvements that could potentially address these:
1) Link "facts" in summaries to lines
2) Expire information after a set time.
3) Turn the backend into a knowledge graph structure
Would love to chat with anyone intrigued in building in this direction!
12
1
64
something to consider when assembling evals for any natural language-driven system:
users WILL ask your system to describe itself & how it works. and it will seem really dumb if it can't.
5
1
66
after thinking about this more i'm also pretty certain an expert LLM answer doesn't solve this for me. i don't want to flatten / take the average of human expertise on a given topic. it feels sad. i want to find the wise human and understand the nuance / story behind their POV
early yahoo search was a list of lists of sources by topic
google revolutionized all that with pagerank
now i just want a list of lists of expertly curated sources again
9
2
89
14
4
62
an omnipresent fact of engineering life, 10 years in
timezones are a pain
6
0
65
i've been surprised how under the radar pgvector has been
don't sleep on this. appreciate
@render
for turning this around in <1 week after some impassioned customer pleas
pgvector is now available for Render PostGreSQL databases!
pgvector is a PostGreSQL extension that lets you harness the power of AI and build semantic search into your apps on Render 🎉
3
8
46
10
2
63
you should be able to build an Assistant in the openai playground by connecting the interfaces of 3rd party integrations with defined tool behavior
e.g. pull in Slack as an integration, search functions to find "send message to channel" -> add as tool to your assistant with a
8
2
61
i think we need a term other than "plugin" for tool use by LLMs. i don't think most people understand what a plugin is
34
4
61
Just learned that the computer onboard Apollo 11, the first rocket that made it to the moon and back, had 2 kilobytes of RAM. 2 KB!!!!!
7
0
61