Varun Mohan
@_mohansolo
Followers
768
Following
123
Media
8
Statuses
538
Explore trending content on Musk Viewer
Oprah
• 624688 Tweets
Papa
• 270757 Tweets
Jill Biden
• 118526 Tweets
Nebraska
• 113735 Tweets
Máximo
• 30703 Tweets
対象作品
• 27438 Tweets
Vaticano
• 26518 Tweets
双葉社フェア開催
• 24546 Tweets
Concord
• 23342 Tweets
Axel
• 19677 Tweets
HAPPY MOMENTS WITH BAMBAM
• 17357 Tweets
Stanford
• 15197 Tweets
Andrade
• 14735 Tweets
Peterson
• 10642 Tweets
Pinned Tweet
Still day 0 in our journey to transform software development. Expect a lot more to come!
Codeium has officially joined the unicorn club! 🦄
Today, we’re announcing $150M in Series C funding, hitting a $1.25B valuation. Special shoutout to our investors
@generalcatalyst
,
@kleinerperkins
& Greenoaks.
More in TechCrunch:
39
58
338
6
3
59
We raised some money from some great folks and am excited to more closely work with
@LM_Braswell
.
We're solving a large and ambitious problem and we're only getting started. If this sounds exciting, we'd love to work with you and are hiring across the board.
🚀 Excited to share Codeium’s new $65M Series B with investment from Kleiner Perkins, General Catalyst, and Green Oaks! 🚀
12
13
142
5
2
61
🚀 AI-powered, natural language based codebase search and synthesis. In IDE. For free. Launching soon on
@codeiumdev
.
Join the waitlist:
1/7
4
6
43
1/ The mark of a good product is meeting the customer wherever the needs are. This is why we are bringing
@codeiumdev
to the browser, the first AI-powered code completion tool to do so. And as always, for free.
1
11
22
Exciting, more to come…
We're excited to share that Anduril is now up and running with Codeium 🎉
For
@anduriltech
, their technology is what sets their company apart making security 🔒 and code quality 💻 their top priority. Thread below 🧵
Read the full case study ➡️
5
1
48
0
0
23
@abacaj
Pretty much every LLM scaling paper suggests more tokens is better. The chinchilla paper even gives a formula for the relation between model size, tokens, and perplexity.
1
0
19
It’s unclear how much compute went into post training for o1 but if it’s not pretraining scale, this will probably mean the capability will become pervasive on open models soon. There would no longer be a an ever growing compute barrier to build models that can reason.
2
0
19
Big milestone!
Big Milestone from the Codeium Community: We've crossed 1M extension downloads! 📈🎉
4
6
86
1
0
18
Next up, 1 million
@codeiumdev
has surpassed 100k downloads on our
#vscode
extension! We remain committed to bringing our users the best AI Developer platform for code completions, chat and search.
3
6
32
0
1
16
@bernhardsson
I think they probably don't understand that Google is at the cutting edge of the field but at the same time kind of right that Google will be slow to push out a product that will disrupt their main business.
0
0
14
Proud of the work the team has done over the past year and excited for the future!
We are excited to announce our $25 million Series A financing, led by Greenoaks with participation from Founders Fund 🎉
1
1
23
1
1
13
If Yale could spend half of the intellectual energy trying to find ways to not let Asians in on actual education, it could be one of the best engineering schools.
0
0
12
Databricks has me working for them after being an intern 6 years ago
Next up on the Codeium integration train:
@databricks
notebooks!
More and more coding is happening in web notebooks (
@ProjectJupyter
,
@GoogleColab
, etc) and they now all have FREE AI-powered code assistance with Codeium 😀
🔗read more:
1
0
7
0
0
11
Cool to share some of the step function improvements we are building to graduate from the Copilot phase of AI software assistants.
1/9 Dream Bigger. Empowering you to do work bigger, better, and faster than before. That’s the Codeium mission, and today we are launching products and plans that take us one step further:
2
1
26
0
0
12
Glad to be represented as one of the preeminent AI startups. We're building a world class AI engineering and go-to-market team and if that sounds exciting, join us -
0
0
12
Huge set of changes we’ve made to improve the chat experience as a whole that spans more than just improving the model. Our new retrieval accuracy is a substantial improvement over embeddings which is incredibly helpful for larger codebases!
1/9 We are rolling out a better chat experience than ChatGPT for software developers. Chat has always been an important part of the Codeium experience, but now it is clearly the best chat experience for anyone writing code.
6
7
84
0
0
11
@spolu
The feed forward layer does way more flops so it dominates the time spent. The attention in and of itself actually doesn’t do that many flops and is memory bound without things like flash attention
1
0
11
@typedfemale
It's pretty fun to watch VCs who thought crypto was the future giving advice on market economics of LLMs and what startups in the space succeed
0
0
10
Super impressive result from
@AI21Labs
. Their model has extremely good eval for its size, beating out Cohere's model which is 3 times bigger. Would love to see how it compares to other open models at similar sizes (Flan-T5) and if they did anything special.
Summary:
- New AI21 model (J1-Grande v2 beta) is quite strong despite only having 17B parameters (3-30x smaller than its peers).
- New Cohere model (52B) also improves over the previous version, passing OPT (175B).
- New OpenAI model (text-davinci-003) requires more discussion...
1
3
32
0
2
11
@tszzl
Is there a capability in the base model you’ve found that the rlhfd one just doesn’t have that’s been super helpful?
1
0
9
My dad would be proud
🎉 Welcome to the Codeium party
#emacs
! We are committed to bringing the power of generative AI to devs _where they dev_.
Recently, we've launched on
#vim
,
@ProjectJupyter
&
@GoogleColab
notebooks, Chrome, and now Emacs.
What's next? 👀
1
0
13
0
0
9
@BucknSF
Data consumption will also go down as companies realize they don’t need to query a table every minute for a dashboard no one looks at
1
0
9
@naklecha
This is super cool. We actually support this for for free without OpenAI using
@codeiumdev
-
2
1
9
Don’t know where to find new cool AI tools? Can’t keep up with the explosion of the gen AI scene? My picks for best AI product aggregators:
Details and links in thread 👇
1
0
8
This is cool but this guy needs to admit
@Tesla
FSD is actually crazy good now. He is losing all his credibility hyping up worse autonomy programs, what did
@elonmusk
do to this man
Congrats to
@MercedesBenz
for being the first automaker in the USA to sell Level 3 autonomous cars that don’t require drivers to watch the road 😉
622
171
3K
1
0
8
🚀🚀🚀 It's been pretty awesome to see how valuable Codeium has been to users so far!
More than a thousand devs started using Codeium in the last week alone 📈 We're just getting started - don't miss out on free AI codegen at !
0
2
10
1
1
8
@amanda_robs
Thanks for having me on! This was a fantastic conversation.
NEW Open Source Startup Podcast featuring AI code companion
@codeiumdev
Founder
@_mohansolo
🎙️
❔How do they compete with a behemoth like CoPilot
❔How did they build a community of 30K
❔How does their infrastructure enable massive scale
Find out👇
1
6
19
1
2
8
This guy codes
🔥 100 days of coding in a row 🔥
I've always been a big fan of streaks. In high school and college, I managed to keep a nearly 2 year GitHub commit streak working on personal projects, schoolwork, and side businesses.
Excited to keep building 🚀
5
1
30
1
0
8
Great to hear
@kevinhou22
provide insight on development at
@codeiumdev
. He’s behind many of the critical product decisions we’ve made!
Pretty cool hearing my voice at 2x speed on my first ever podcast appearance on
@SyntaxFM
!
- How do you build an AI coding assistant? 🧠
- What context does
@CodeiumDev
's LLM use?
- How & why do we give a 100k+ DAU AI product away for free?
Listen ⬇️
5
7
38
0
0
7
6/ And finally, follow
@codeiumdev
for more content on the future of AI-assisted coding. We have lots of exciting never-seen-before AI capabilities coming really soon. Happy coding!
0
0
6
Excited to see how folks use Chat!
We’re excited to introduce Codeium Chat!
It's available today for VS Code in beta.
Find out how to get free access 🧵
8
21
70
0
1
7
@Si_Boehm
Numbers are a bit off. 30/13 is 2.3x not 3x. V100 fp16 was actually 130 for non PCie. H100 with fp8 is a > 6x and atleast a 3x otherwise. Training improvements are largely driven by compute increase so if fp8 works, it's a much larger jump than V100 on price / performance
0
0
7
@typedfemale
Writing a compiler is bad. Means people might write a backend for ROCm and that would be no bueno
1
0
7
Glad
@_anshulr
and I were able to join
@changelog
and talk about the generative AI applied to software and our journey at
@codeiumdev
. Thank you for hosting
@chrisbenson
-
0
0
6
@typedfemale
TPUs and the softmax in self-attention make me sad. Nvidia does the right thing and lets you fuse the reduction into the matmuls
1
0
6
Excited to roll out our offering to folks with self-hosted and finetuning requirements!
Introducing Codeium for Enterprises 🚀
🔒 On-prem or in your Virtual Private Cloud (VPC)
🛠️ Fine-tuned models on your existing repo entirely self-hosted
🧠 Full suite of AI acceleration for coding including autocomplete, search, and more
Learn more 🧵
1
1
16
1
2
6
@typedfemale
MIT had CPUs that were 10 years old when I was there for the systems group - I don’t think this will happen
0
0
6
Excited to share Codeium (), a free forever version of Copilot! We think this is still day 1 🚀
Big updates!
⌨ VSCode and JetBrains support
⚡ 2 minute installation process
📈 More languages supported, better quality suggestions, even faster speeds
Head over to to see everything Codeium can help with and get started today for free!
0
3
14
1
0
6
The moment a metric can be juiced, it will be juiced. Acceptance rate for instance is not a great metric if you give fewer suggestions at higher latency (that are also shorter on average).
Don't know what benchmarks and numbers to trust when evaluating AI tools? In our latest blog, we take a deeper look at these metrics for AI coding assistants and describe how we approach evaluation at Codeium!
🔗
0
0
9
0
0
6
This just makes me want to code more...
Announcing Codeium Wrapped 2023! You can check out your coding patterns with Codeium at:
0
2
58
1
1
6
There's a false understanding that self-hosting is hard and costly because most vendors think about this as an afterthought after developing a SaaS product. We didn't!
Codeium can be self-hosted and run in your VPC or on your hardware. Unlike other on-prem software, we are:
- Easy to setup, update, & maintain ✅
- Flexible hardware for all budgets 💰
- Adapts to fit your unique infra 🛠️
Article with the details ⬇️
0
5
37
1
0
6
Glad we could share more about how we think of differentiating Codeium from a product perspective and why hundreds of thousands of developers use us.
0
1
6
2/ How does it work? Our embedding model will compute and index embeddings over your codebase ON YOUR MACHINE. Your code or derived embeddings never leave your machine.
1
0
4
@bernhardsson
I think it’s dependent on the space. If everything below the product is a commodity, that’s true. Products like Lensa or Midjourney are dominant because they actually built good tech in a nascent space (and better than the infra layer cos)
0
0
5
Early days but we’re just getting started 🚀
On the recommendation of
@swyx
, I turned off my Github Copilot and tried
@codeiumdev
. I can confirm Codeium is working as well as Copilot. Impressive feat!
2
2
34
0
0
6
We are still committed to delivering best in class products free forever for individual users! We’ll talk more about all the things we have to do to make this a reality (maybe a podcast is in order then
@swyx
@FanaHOVA
)
Running LLM's at scale for hundreds of thousands of users is not easy. At Codeium, we are in the business of doing hard things in order to bring the best AI-powered dev tools to our users...at $0 / user / month
We are proud to continue offering 100% free, forever autocomplete &
4
5
41
0
0
6
@swyx
This is too true. I recently tried to be smart and use RocksDB because it gave me some nice snapshot isolation guarantees and then later decided to use SQLite the moment I realized I needed a metadata column
2
1
6
I want to build a company that uses JIRA as a realtime database. You can comment when an update is incorrect and roll them back appropriately. Change my mind.
1
0
5
GPT-4 is good
ChatGPT is great
AutoGPT is the future
Let's dig into this 🧵
3
0
5
@buckymoore
@AIatMeta
@AnthropicAI
@OpenAI
I think the answer is not much prevents switching to llama. As a default model, GPT-5 may become the most powerful but I can see a world in which most model inference is llama based and (while meta is catching up), the most expensive and powerful general model is openai based - I
0
0
5
Great to see
@kevinchen
getting his dues
This
@YouTube
channel owned by one Mr. Chen (
@kevinchen
) is likely one of the most underrated Autonomous Vehicle channels that I know of.
For example here is a great commentary by him of a
@Waymo
ride: .
Definitely worth subscribing to; I did, at least.
1
2
12
1
0
5
@abacaj
Actually working out the chinchilla formula for a 7B, for it to match performance of the 70B - You would need 40T tokens (so ~3x the compute as training the 70B).
0
0
5
🎧Excited to release our first AI Infra podcast!
With
@_mohansolo
of
@codeiumdev
, we discuss:
- How to cut inference costs by 90%+ (like
@OpenAI
just did)
- Copilot for X (Law - Harvey)
- Scaling Laws vs Serving
- Key needs for LLMOps
Who else should we talk to? pls request!
3
6
38
0
0
5
@typedfemale
Why do you say that? 312Tflops to 1Pflop bfloat16 is already a bigger jump than V100 to A100 (125 -> 312). If fp8 works (which it seems like it will) for a nontrivial chunk, the win is going to be well over 3x
2
0
5
3. theresanaiforthat
Largest catalog and finer grained “tasks” (as opposed to categories). The timeline view gives a rough sense of when these tools have popped up.
🔗
1
0
5
7/ This is just the next step in our mission at
@codeiumdev
to build a one-stop AI-powered code acceleration toolkit. Oh, and in case you missed it, it’s free.
Join the waitlist:
0
1
4
Excited to share how we are pushing the needle on enterprise-grade AI developer tools. Context is great, but there are limits. We show how we outperform Copilot with context with a locally fine-tuned model without any context!
Many of our enterprise customers see large suggestion quality improvements from our automated, locally-executed personalization features! Here's some proof that suggestions do get better:
🔗
1
9
28
0
0
5
4/ No other AI-powered code acceleration toolkit works on arbitrary web editors. And
@codeiumdev
is one of the best in quality and latency anyways:
1
0
4
Huge time saver! Just the beginning for Termium.
k for me now always autocompletes to kubectl get pods...
🚀 Excited to Launch Codeium in your Terminal: Termium 🚀
With Termium, you get unlimited autocomplete in your Terminal!
12
20
126
0
0
5
@Tim_Dettmers
The network requirements actually should just get better since you can also store your weights as fp8 which is beneficial for fsdp. Networking requirements scale sublinearly with compute so it should actually help more than it hurts.
1
0
3
@typedfemale
Agree on sparse transformer (do they still do this?) but UL2 mainly just says to use span corruption on pretraining, which just seems more general than stuff like FIM.
1
0
4
0
0
4
Big milestone for us on security guarantees on all fronts - self hosted and cloud managed!
🔒 Codeium is SOC 2 Type 2 Certified🔒
While other managed tools like GitHub Copilot do not even have SOC 2 Type 1 certification yet, we have built the most secure hosted code assistant.
Read more:
0
2
53
0
1
4
@rauchg
@ClickHouseDB
We made a playground for our product (a Copilot alternative)
@codeiumdev
- that lets user have the same interactive experience using our product in an IDE and it's been quite helpful.
0
0
4
@DynamicWebPaige
@GoogleColab
@OpenAI
You should also give us a try out at
@codeiumdev
we have a free chrome extension (no OpenAI API key required) that does code completion in Colab and Jupyter notebooks.
0
0
4
@DrJimFan
Thanks for the summary. 20M seems quite steep to train this model even on A100s, where does that come from?
1
0
3
Glad we could share some of the internals of building a Copilot-like product. Also happy to get a stamp of approval from
@swyx
😃
🆕 What Building "Copilot for X" Really Takes
Everyone's talking about "Copilot for X", but few have actually built one.
Presenting my first guest-post from
@_anshulr
, on building
@codeiumdev
! The first ever roadmap to a "generative AI moat" I've seen.
3
27
112
0
0
4
Thanks a lot
@_seanyneutron
I honestly have 0 clue how to read vimscript
Big thanks to
@_seanyneutron
for writing basically all of the plugin! We welcome further contributions!
0
0
4
0
0
4
Very glad to be working with the team
@greenoakscap
and Neil. It was very easy to tweet and offer thoughts and prayers last week but takes much more courage to actually do something.
1
1
4
@_seanyneutron
- the silent hero of all open source projects
It's finally here...
⚡️Javascript/TypeScript support for LangChain⚡️
BIG thank you to everyone in the community who contributed to this, especially
@_seanyneutron
and
@nfcampos
❤️
GitHub:
Blog:
LOTS to discuss, lets 🧵
21
79
487
0
1
3
@codeiumdev
has lots of Princeton alums and looking for many more :)
I'll be in NYC and Princeton next week to meet with the best and brightest
@Princeton
and
@Columbia
computer science students. Anyone I should meet? DMs open.
@CodeiumDev
is hiring like crazy, especially new grads. I can confirm there's never been a more exciting time to join!
1
3
31
0
0
4
5/ We are currently enabling this extension on a whitelist-basis for websites. We will be keeping the Chrome download page up to date with all compatible websites:
1
0
3
@dylan522p
Cool article, I think the cost per inference is probably a factor of 3 lower. Basically it takes 7 A100 seconds assuming 50% util per chat and I think per hour pricing is probably $0.6/hour (given preemptible on GCP is $1/hour), which is like 11 cents a chat.
1
0
4
@code_star
200k for you but for them it is 1.85x larger model trained 1.4x more tokens at 2x the price, assuming they chose on-demands instead of $2/hour, which comes out to 5x. Feels like the efficiency is roughly the same.
2
0
4
3/ Now, in less than 2 minutes, any developer can install the free
@codeiumdev
Chrome extension and get the same in-IDE coding superpowers in the browser:
1
0
3
I had a great time talking about the AI in the future of engineering with
@Dpbrinkm
. Thanks for having me on!
Hey, just dropped the latest MLOps Community podcast episode and it's a real game-changer!
@_mohansolo
, the CEO of
@exafunction
and
@codeiumdev
, spilled the beans on Building the Future of AI in Software Development.
1
1
4
0
0
4
Big changes! Chat history has been a long requested feature
🚨 Another huge week of Codeium updates: multi-conversation Chat history, increased codebase awareness, public profiles, Codeium Command ✨, and much more
Update to our latest release on VS Code and Jetbrains today!
100% free, forever AI Dev Tools ➡️
6
7
32
0
1
4
@typedfemale
I think good implementations only compute the lower triangle (like flash attention). It does feel dumb to not compute the whole thing when attention is a small number of total flops
0
0
3
Am I misreading the eval here? This seems like no improvement compared to the base CodeGen model. Feels like OpenAI's edit endpoint is still substantially better then.
1
0
3
6/ We will keep iterating on the embeddings to get closer to the exact answers, but we are at a point where we are confident it adds value today.
1
0
2