Itamar Friedman
@itamar_mar
Followers
5,283
Following
417
Media
170
Statuses
917
Excited about the future of intelligent software development. CEO & co-founder @CodiumAI
TLV
Joined October 2013
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
スプリンターズS
• 88616 Tweets
LIVE YOUR DREAM APO
• 67519 Tweets
かるびCR
• 55250 Tweets
CRかるび
• 53184 Tweets
WIN AT TODAY SHOW
• 51089 Tweets
REBECCA SUPPORTING ACTRESS
• 42047 Tweets
YINWAR x VIF x TOPS
• 37179 Tweets
ナムラクレア
• 25874 Tweets
ショアキーパー
• 21029 Tweets
西村騎手
• 20926 Tweets
ジャック
• 20438 Tweets
YIBO PLAY WITH LACOSTE
• 15910 Tweets
Pinned Tweet
1/🚀 Introducing PR-Agent Chrome Extension,
allowing any developer to chat with AI directly on pull requests in GitHub,
powered by top code models like Claude 3.5 Sonnet and GPT4o!
8
19
76
🚀 Introducing AlphaCodium -
A first-of-its-kind open-source code generation tool that surpasses most human competitors in code contests ⭐️
Inspired by DeepMind's AlphaCode❤️🔥, but beats it (judge by yourself!)
1/
20
175
918
🚀 Introducing Cover-Agent 🧪
An open-source tool that includes a reimplementation of Meta's TestGen-LLM for automatically enhancing test suites.
Manager: "We must improve old test suites for better code coverage. Can you handle it?"
Me: "Sure, my favorite task... (Not!) 🤷♂️"
20
207
929
@svpino
Hey 👋, one of the Cover-Agent creators here.
I've recorded 5 minutes video explaining and reviewing TestGen-LLM and Cover-Agent:
🚀 Introducing Cover-Agent 🧪
An open-source tool that includes a reimplementation of Meta's TestGen-LLM for automatically enhancing test suites.
Manager: "We must improve old test suites for better code coverage. Can you handle it?"
Me: "Sure, my favorite task... (Not!) 🤷♂️"
20
207
929
8
33
322
I’m a big believer in AI coding agents.
But I don’t think the way to get to fully autonomous AI software engineers is by jumping to “self-driving” agents.
@CodiumAI
just released a different type of agent. It’s embedded in the IDE and works in tandem with you as you code.
1/
7
28
248
Super excited to announce that we’ve just launched
@CodiumAI
, the product and company, and raised $11M seed round 🚀🚀
CodiumAI generates meaningful tests for busy devs.
Check it out:
Here is my announcement:
1/
13
32
222
@karpathy
@Kyrannio
One of the makers here 👋
I've recorded a 5min video explaining AlphaCodium in high level:
We estimated that we spent more than 95% of our research time on flow engineering rather prompt engineering
🚀 Introducing AlphaCodium -
A first-of-its-kind open-source code generation tool that surpasses most human competitors in code contests ⭐️
Inspired by DeepMind's AlphaCode❤️🔥, but beats it (judge by yourself!)
1/
20
175
918
13
9
221
my main takeaways from
@karpathy
:
1> agents are expected to have a huge real impact on our life... Like autonomous cars promise.
2> but: similar to AV, people overestimate the difficulty of building a real agents-empowerd products, not just a demo.
Karpathy says it might take a
5
21
201
6 technological advancements are likely to increase the (justifiable?) hype around Generative AI and Large-Language-Models (LLMs)
In three years' time, most of the limitations of today's LLMs will be eliminated
Arguments in blog and thread 👇
6
45
179
🚀 introducing pr-agent - get PR analysis and suggestions inside your GitHub Pull Request
you can either try the open-source:
or simply summon
@CodiumAI
-Agent on any GitHub public PR 🤯
let's dive into what sets pr-agent apart:
6
36
164
@svpino
I've recorded a 5 min video explaining AlphaCodium
🚀 Introducing AlphaCodium -
A first-of-its-kind open-source code generation tool that surpasses most human competitors in code contests ⭐️
Inspired by DeepMind's AlphaCode❤️🔥, but beats it (judge by yourself!)
1/
20
175
918
5
20
164
"Agent GPT/LLM" refers to an autonomous AI-based program capable of interacting with its configurable environment & tools to complete requested tasks
Noteworthy open-source projects:
• AutoGPT by
@SigGravitas
- buzzy
•
@LangChainAI
's Agent - well established & versatile
👇
9
23
157
pick your AI programmer friend 🤖:
'gpt-engineer' -
@antonosika
'smol-dev' -
@swyx
'AutoGPT' -
@SigGravitas
'Metamon' -
@yoheinakajima
they commonly "work" this way:
▸ You give a first set of instructions
▸ AI asks clarifying questions, generates spec, writes code, 🔁
👇
10
18
150
2024 software development:
❶ Developer writes specs (w/ auto-complete),
❷ AI Agents generate code & tests,
developer reviews & edits,
❸ AI Agents deploy,
developer reviews & approves
💫a proof-of-concept with
#AutoGPT
(by
@SigGravitas
) and
@CodiumAI
(w/ TestGPT):
4
25
138
AlphaCodium is open sourced⭐️
It includes the complete AlphaCodium code, fully reproducible, and scripts to apply it to Codeforces problems
Let's accelerate the development of code generation tools that produce code that actually works
4/
2
19
131
ChatGPT Code Interpreter plugin is a game changer 💫
🤖 Agents and LLMs tooling frameworks also have code execution capabilities.
e.g.
@LangChainAI
is equipped with the Python REPL tool
Now,
@CodiumAI
released this "run tests" then "reflect & fix" 👇
We're thrilled to announce the release of CodiumAI 0.5.20 for VSCode 🚀
Get ready for some fresh & exciting features 🥁
★ CodiumAI runs tests and fixes them if needed 💫 "Reflect & fix"
★ CodiumAI can think harder on specific tests and improve them 🧐 "Reflect and regenerate"
1
5
48
5
17
103
AutoGPT roadmap is out 💫
@SigGravitas
(the inventor of
@Auto_GPT
) shared details in the below video (start at ~0:07:00)
★ Making AutoGPT accessible to everyone via mobile & web apps
★ Development principles: Challenge-driven-development & Plugins
More details and takeaways👇
7
27
96
AlphaCodium X posts curated🧵
Interesting prototyping, explanations, and insights.
1/ DSPy
DSPy lets you prototype LLM Programs like AlphaCodium in 2 minutes! 🧩🔥
12
63
404
1
9
97
🚀Two game-changing dev tools were unveiled today.
Say goodbye to fully manual & frustrating unit-test creation and hello to AI-assisted & fun test generation.
@GitHubNext
released Test Pilot for typescript/javascript developers, and ... 1/🧵
Take your test pilot for a spin: GitHub Copilot Labs now comes with a test generator, that creates and refines tests!
@GitHubNext
63
585
4K
3
12
81
Code Q&A using RAG for large code bases has unique challenges
We are now sharing how we used
>
@llama_index
,
> static analysis,
> advanced chunking paradigm,
to deliver a working solution
1/
6
12
109
@gdibner
Re: "opportunity for LLMs ... is HUGE, but ... far smaller than what most people believe (because ... inflated relative to reality)."
Amara's law applies here:
We tend to overestimate the effect of a technology in the short-run and underestimate the effect in the long-run!
1
4
67
AlphaCodium:
‣ Open-source: ⭐️
‣ Paper:
‣ Blog:
‣ Discord:
Huge credits to the main author and maker,
@talrid23
👏
6/6
3
10
66
6 Best practices are key factors in AlphaCodium's code-oriented flow
Try them out when you use LLMs for code-generation tasks!
3/
2
10
64
GPT-4 is out! And it is stupendous!
𝗕𝘂𝘁 𝗶𝘁 𝘀𝗲𝗲𝗺𝘀 𝘁𝗵𝗮𝘁 𝗶𝘁 𝗶𝘀𝗻'𝘁 𝘃𝗲𝗿𝘆 𝘀𝗸𝗶𝗹𝗹𝗲𝗱 𝗮𝘁 𝗽𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴.
What? Why? How can it be?
It does so well on many other exams!
And the coding demos are so wonderful!
Let's discuss 👇
3
4
57
New code models are continuously released, with Codestral being the latest.
They are mostly compared on the HumanEval benchmark. Open-weight code models roughly reach similar results and perform quite poorly vs. closed models.
To reach high-quality results with open code models
🚀 How good actually is the new 'Codestral' model? 🤔 And which model should you choose to fine-tune for your specific code task?
Discover the new PR-Agent fine-tuning benchmark!
It methodically compares various open-source models based on their fine-tuning capabilities. Check
3
0
5
1
2
53
Very much agree!
Here is a practical example:
Let's say you are starting a new project, that involves choosing a database stack and schema.
Three ways to do it:
1> Think really hard, and make the perfect selection.
Wrong. You just can't. The right solution for today will be
Sunday rant.
For software engineering, my sense is that the phrase “premature optimization is the root of all evil” has massively backfired. Its from a book on data structures and mainly tried to dissuade people from prematurely write things in assembler. But the point was to
225
843
6K
6
0
51
"AI [Developer] Teammates" will emerge eventually!
Dev tools that successfully land net-new capabilities and have strong connections to other key parts of the software-development-life-cycle will have the potential to expand and become a fully-fledged developer teammate
🤔🧵👇
2
6
48
Human-designed neural networks are [still] at least as efficient as those designed by Neural Architecture Search. Perhaps we need to rethink our NAS objectives.
e.g. see our latest TResNet:
• Paper:
• In
@wightmanr
awesome repo:
3
10
49
@amasad
❤️🩹 thank you for sharing 💚
Today Israeli Arabs live in Haifa and many other cities all around Israel.
I hope you can come for a visit and see how Israel grew up becoming a multi-religion & multi-national country, e.g. 20% of the population is Arab (still a lot to improve
12
0
45
Original TestGen-LLM paper:
Cover-Agent open-source that reimplements TestGen-LLM by Meta:
2
9
47
Are ChatGPT or GPT-4 knowledgeable systems, intelligent systems, or both?
After-all, ChatGPT recently past some challenging medical, programming and other MBA exams
So, they are definitely knowledgeable in various fields, don’t you already agree?
But are they intelligent?
1/
4
5
46
AlphaCodium differs from AlphaCode in 3 ways:
1‣ AlphaCodium works with any leading code generation model.
Hence, it's a generic solution, while AlphaCode2 exploits the Gemini-Pro model that was explicitly fine-tuned for the Codeforces competition 😏
2.1/
1
3
42
"You see here I have a sad face ;-( , since tokenization is my least favorite part of working with LLMS, but unfortunately, this is necessary to understand"
There are many reasons why people would consider
@karpathy
as a 🐐 AI Engineer.
Tackling essential aspects, even if they
New (2h13m 😅) lecture: "Let's build the GPT Tokenizer"
Tokenizers are a completely separate stage of the LLM pipeline: they have their own training set, training algorithm (Byte Pair Encoding), and after training implement two functions: encode() from strings to tokens, and
379
2K
14K
2
3
42
2‣ AlphaCodium is an open-source, available tool, while AlphaCode is a research paper
3‣ AlphaCodium strongly emphasizes code testing and code analysis methods and flows, which are specialties of
@CodiumAI
/2.2
1
1
41
Just 400 years ago, most people believed that the sun revolves around the Earth.
I find this incredible and almost impossible to imagine.
What fundamental truth that is contrary to our current beliefs will be obvious to next century people?
Consciousness is one of my two
8
3
40
It is fun and satisfying to see a developer (let alone
@swyx
) wearing your company's swag 😎🆒
6
2
38
ImageNet-21K is very likely [much] better for pre-training vs ImageNet-1K. Yet, it isn't straightforward to use it.
Check out this work by
@talrid23
et al.
This newly suggested processed ImageNet21K is now an official part of
1/2
1
7
38
@hwchase17
Agents in production:
• 70% Code & tooling - Flow engineering.
• 20% AI - Obviously essential, can't do agents without the models.
• 10% Human Intelligence - Indispensable, and thus agents are bound by the UX/UI!
Cc:
@talrid23
, the first author of AlphaCodium
3
5
36
What a great day!
It started with a workshop we delivered to a group of developers from one of our fantastic clients,
@HiBob_HR
Their questions and feedback were excellent.
One of their suggestions really sunk in, and we decided to develop not one, but two features, both of
1
2
38
"How Smart Is a Rock?
To appreciate the feasibility of computing with no energy and no heat, consider the computation that takes place in an ordinary rock.
Although it may appear that nothing much is going on inside a rock, the approximately 1025 (ten trillion trillion) atoms
"I grew up implicitly thinking that intelligence was this, like really special human thing and kind of somewhat magical. And I now think that it's sort of a fundamental property of matter..."
@sama
8
23
160
2
7
36
Agent Weekend Test
5
9
34
#RAG
can really boost code search and generation 💻✨
But,
when dealing with large enterprise codebases, a reliable and accurate solution comes with its own set of challenges.
This system, shown in the diagram, processes code files, breaks them into meaningful chunks, generates
3
4
120
Are AI “Agents” simply a rebranding of “Chains”?
Can Chains be used by Agents or vice-versa?
Actually,
Agents can use Chains as tools,
and Chains can chain Agents.
Let’s discuss 👇 (including a short list of Agents)
7
3
34
I've noticed that my working time on airplanes is very efficient, maybe even the most efficient.
I'm thinking that perhaps we should invest in a very small booth for the office where you need to buckle up, close the door, and can't leave for a few hours.
4
2
33
@emollick
My suggestions, before claiming that GPT-4 has been beaten:
Let's wait a few days and see some interesting benchmarks done by interesting people.
We are living in extraordinary times!
But why does this report say 67% for GPT-4 on HumanEval when people have demonstrated 84% or higher multiple times?
It makes me suspicious.
4
3
22
2
1
33
AlphaCodium flow is a practical one
Compared to AlphaCode 1, it requires 4 magnitudes less of LLM calls
Compared to AlphaCode 2, it reaches on-par results with the same amount of LLM calls,
but AlphaCodium does not require fine-tuning a model!
Generalization is important
5/
1
5
34
@karpathy
, just highlighting an IMPORTANT point:
Besides improving accuracy, "flow engineering" also really helps with reducing variance‼️
Quoting from GPT-4 Technical Report: "roughly 50% of simulations have 0 problems solved"
-- This doesn't happen with AlphaCodium::GPT-4
Prompt engineering (or rather "Flow engineering") intensifies for code generation. Great reading and a reminder of how much alpha there is (pass
@5
19% to 44%) in moving from a naive prompt:answer paradigm to a "flow" paradigm, where the answer is constructed iteratively.
126
550
3K
2
1
31
I just wanted to grab lunch and bumped into the amazing teams
@gpt_engineer
@sanalabs
@AgentOpsAI
@CodiumAI
AI is happening in SF?
6
0
31
ChatGPT amazed the world and was the fastest product ever to reach 100 million users
With a simple and intuitive interface, ChatGPT enabled literally anyone to converse with a GPT-empowered chatbot for free
It also enabled everyone to explore the current limitations of LLMs
2/N
1
6
31
As a developer 👩💻, here is how you can improve (or ruin 🤬) your relationship with your boss:
3
5
31
@svpino
I've recorded a 2min video showcasing the PR-Agent highlighting issues in a PR from yesterday!
We are working hard to enable PR-Agent to integrate seamlessly with any Git platform
and catch issues before they hit production!
Can we exploit AI to reduce software outages?
PR-Agent (open-source) by
@CodiumAI
might be the way to go
1
1
16
2
5
70
|￣￣￣￣￣￣￣￣￣￣￣￣￣|
"Everything That Can Be Done with AI, Will Be"
|＿＿＿＿＿＿＿＿＿＿＿＿＿|
\ (•◡•) /
\ /
——
| |
|_ |_
The origin:
2
5
29
I strongly agree.
AI's potential to assist in areas where humans face significant challenges, such as code verification and bug finding, could indeed be game-changing.
Regarding *formal* verification:
It's a rigorous process aimed at proving or disproving the correctness of a
One application of AI that I am excited about is AI-assisted formal verification of code and bug finding.
Right now ethereum's biggest technical risk probably is bugs in code, and anything that could significantly change the game on that would be amazing.
3K
3K
15K
1
4
24
I sat down with
@petercohan
to discuss the development of AI as a game changer for improving Enterprise workflows for software developers… hint: the key is
#FlowEngineering
!
The recognition of the importance of the flow engineering is growing, and
4
3
27
CodiumAI focuses on Code Integrity
Capabilities & features are all about challenging & verifying code correctness, so developers can code fast with confidence
Don't automatically trust code suggestions from code generators like Copilot
Check your code. Own your code 🧑💻
We are on a roll!
CodiumAI releases v0.5.23 for VSCode 🚀
★ CodiumAI generates tests & runs them. If a test fails, CodiumAI suggests a test fix or an application code fix 💫
Try it out!
4
7
21
1
1
23
"Self-healing code is the future of software development" by
@benpopper
(
@StackOverflow
)
To self-heal, the coding assistant/agent must be capable of creating tests, running them, reading errors, and r/w/converse about code specification
6
7
25
Many AI coding assistants, like Copilot, heavily rely on existing code for context, which could pose risks.
Example ❶
In the video,
@snyksec
demonstrates that if there is a vulnerable code snippet in a neighboring tab, Copilot still incorporates it into its context and can
4
1
23
@hwchase17
@jordnb
What about -
Agent shows next step without executing, press tab to execute,
or edit before executing.
Think code-auto-complete-style, but in an agent flow.
2
3
23
We are living in extraordinary times!
But why does this report say 67% for GPT-4 on HumanEval when people have demonstrated 84% or higher multiple times?
It makes me suspicious.
Today, we're announcing Claude 3, our next generation of AI models.
The three state-of-the-art models—Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku—set new industry benchmarks across reasoning, math, coding, multilingual understanding, and vision.
570
2K
10K
4
3
22
@swyx
AI Engineering will shift from prompt engineering to flow engineering
I'm one of the makers of the paper & work (AlphaCodium) Karpathy is quoting.
We estimated that 95% of our research work was on the flow design and experimentation !!
5min explanation:
🚀 Introducing AlphaCodium -
A first-of-its-kind open-source code generation tool that surpasses most human competitors in code contests ⭐️
Inspired by DeepMind's AlphaCode❤️🔥, but beats it (judge by yourself!)
1/
20
175
918
0
2
22
@yoheinakajima
Coders, like gardeners,
won't create the initial plants (code) but instead nurture and refine them according to their desire
Coders, like orchestra conductor,
won't play the instruments, but instead shape & orchestrate the performance of music generated by the AI
This is fun
3
1
21
Old Coder Guy uses and teases
@CodiumAI
to test its code logic 🤣
He also makes fun of me, but I'm totally fine with that 😀
Writing tests could be a frustrating task, so pouring in some fun and artificial intelligence is the way to go!
Good job!
0
7
21
Exactly a year ago, I tweeted about
@CodiumAI
for the first time, inviting developers to join our closed-alpha program and test our very first VSCode IDE extension.
We then embarked on our journey to empower developers to code, test, and merge with confidence.
Our initial
5
5
22
AI-generated code that enters into your code base is of lower quality than human-generated code, they report in Visual Studio Magazine:
I don’t think it is a problem with the foundation models, but rather how we use them,
it is essentially a UX/UI/product
3
1
21
@svpino
Last week, I developed a new feature in AlphaCodium, and PR-Agent saved me from an embarrassing bug.
I know that rigorous testing should catch all these cases, but having "testless" semantic testing is fantastic.
Here is a link to the pull request:
You
0
2
20
In the dev realm, there are two camps:
1. You don't write tests, since you hate it
2. Or you do write tests, ... and hate it 🤦
AI might actually change this 💡
Impressive features 👇
Big shoutout to
@talrid23
&
@hussam_lawen
👏 on the neat implementation
We get you.
Life gets busy 🧑💻.
You hit submit on that pull request,
only to realize testing and documenting slipped your mind🤦🏽♂️
No stress, PR-Agent has your back ✨
Check out these new features – effortlessly and interactively generate:
🧪 Test ideas and draft implementation
1
2
35
1
2
19
@cohen_eyal4
@CodiumAI
This open-source tool is still a research project. Throughout 2024, we will work diligently to bring this technology to your fingertips, integrating it within the CodiumAI IDE and Git plugins
1
0
20
i wrote about AI Agents & SW 3.0 on March 2022!
do you think that the vision of AI Agents will be realized before 2025?
curious, would you consider `
#ChatGPT
+ Code Interrupter` as an agent?
some selected quotes from the blog 👇
did it age well?
2
1
20
@skies_dev
@code
in the post they write:
.... bring the power of generative AI and GPT-4 throughout the entire developer experience on GitHub...
3
0
20
@DrJimFan
AlphaCode is more of a system than "just" a model.
AlphaCode includes a model fine-tuned specifically for the competition purpose, but the core concept revolves around extensive sampling/generation of various solutions, followed by a "smart" selection process involving
0
1
19
@alighodsi
@databricks
"dolly-v2-12b is not a state-of-the-art generative language model and, though quantitative benchmarking is ongoing, is not designed to perform competitively with more modern model architectures or models subject to larger pretraining corpuses."
Found this
2
1
17
is ChatGPT evolving from a chatbot into an autonomous AI agent?
in other words,
will we be able to request ChatGPT to complete tasks for us autonomously, including making decisions and taking action?
2
4
19
Follow, retweet, like, or reply to get my following tweets - I will be arguing about
#chatgpt
, GPT-3/4.5 & AlphaGo intelligence
A big thank you to the post reviewers!
@mathemagic1an
@talrid23
@GadiZimerman
@NetaBarkay
,
@TranPatrik
,
@mradamjafer
, Amit Mandelbaum,
@mghissassi
🧵
0
3
17
@LangChainAI
When building real world AI empowered systems, we see a shift from prompt engineering to flow (/graph) engineering.
LangGraph and AlphaCodium are a perfect fit 👏
AI tools will help us busy developer to code with on the fly generated tests.
3
1
16
Jamba -
Competes with Mixtral 8*7B and alike.
What's special?
> Apache 2.0 license! Open Weights.
> Mixed: Mamba + Transformer architecture
> Gain in performance long context
> MoE 52B
What next:
> Releasing an instruct model
> Releasing a model with tool use capabilities
Introducing Jamba, our groundbreaking SSM-Transformer open model!
As the first production-grade model based on Mamba architecture, Jamba achieves an unprecedented 3X throughput and fits 140K context on a single GPU.
🥂Meet Jamba
🔨Build on
@huggingface
37
252
1K
0
0
16
I agree with this list, although I believe that evals are fundamental also to AI/ML engineers.
I actually claimed today, when chatting with a friend on this topic, that having experience with planning, creating and using evals might be the most important aspect of AI engineering
3
3
15
@HamelHusain
Database -> files with information
Cloud -> connected computers
Network cables -> conductive wires
Airpods -> Apple's headphones
2
0
17
Shortlist of LLMs' limitations:
1. Content hallucinations
2. Black-box behavior
3. Requires massive training
4. Updating a model is cumbersome
5. Input size is too small
6. Very basic reasoning
7. Not optimizing for a global cause
8. Heavily dependent on prompt engineering
3/N
2
2
17
'/test' your code
Copilot just announced on Copilot Chat GA in a few weeks 🎉
𝐛𝐮𝐭 𝐰𝐡𝐲 𝐰𝐚𝐢𝐭!
@CodiumAI
offers an advanced Chat including an unparalleled '/test' command already today 😃
In the thread below, I compare CodiumAI /test to
@GitHubCopilot
one
1
3
16
According to my rough estimation,
𝐃𝐞𝐯𝐢𝐧's March results would position the tool around 10th place on the SWE-bench (Lite).
Is this estimation accurate?
@cognition_labs
@jyangballin
@_carlosejimenez
1/4
Today we're excited to introduce Devin, the first AI software engineer.
Devin is the new state-of-the-art on the SWE-Bench coding benchmark, has successfully passed practical engineering interviews from leading AI companies, and has even completed real jobs on Upwork.
Devin is
5K
11K
45K
1
0
15
Impressive 👏
-> respect for open sourcing the training and benchmark data, making the model and work fully reproducible and auditable
-> "text embedding model with a 8192 context-length that outperforms OpenAI Ada-002 and text-embedding-3-small on both short and long context
Introducing Nomic Embed - the first fully open long context text embedder to beat OpenAI
- Open source, open weights, open data
- Beats OpenAI text-embeding-3-small and Ada on short and long context benchmarks
- Day 1 integrations with
@langchain
,
@llama
-index,
@MongoDB
38
272
2K
0
2
16
@AviSchiffmann
have you considered that we people don't know exactly what we want?
applications encode product decisions made by their creators, reducing the mental load for the user to think about everything
0
0
15
Can we exploit AI to reduce software outages?
PR-Agent (open-source) by
@CodiumAI
might be the way to go
1
1
16
2/ Karpathy's take on 'flow engineering'
Prompt engineering (or rather "Flow engineering") intensifies for code generation. Great reading and a reminder of how much alpha there is (pass
@5
19% to 44%) in moving from a naive prompt:answer paradigm to a "flow" paradigm, where the answer is constructed iteratively.
126
550
3K
1
1
16
An independent testing agent filters for the best code suggestions to meet the requirements. It understands the context of the code within the application and can run alongside any primary code gen model tool.
It happens to be that this is what we are working on at
@CodiumAI
:)
0
0
16
@gregisenberg
For those asking to learn more about AutoGPT, and about AI Agents in general, maybe this thread could help
Now about ChatGPT, I expect it to have Agent-like capabilities and features very soon, via its plugins
"Agent GPT/LLM" refers to an autonomous AI-based program capable of interacting with its configurable environment & tools to complete requested tasks
Noteworthy open-source projects:
• AutoGPT by
@SigGravitas
- buzzy
•
@LangChainAI
's Agent - well established & versatile
👇
9
23
157
1
1
16
1
0
16
Will we very soon see implementation of Code Generation tools like Copilot but with models that runs surprisingly fast on your MacBook/computer?
"How is LLaMa.cpp possible?"
great post by
@finbarrtimbers
llama.cpp surprised many people (myself included) with how quickly you can run large LLMs on small computers, e.g. 7B runs @ ~16 tok/s on a MacBook. Wait don't you need supercomputers to work
81
739
5K
2
3
15
Meta's TestGen-LLM provides another glimpse into the future of software development.
It offers a compelling use case and results: The system generates test cases that can improve code coverage.
1> It first generates a bunch of tests, then filters out those that don't run, pass,
Meta’s LLM for software testing work is super exciting.
This paper describes Meta’s TestGen-LLM tool, which uses LLMs to automatically improve existing human-written tests. TestGen-LLM verifies that its generated test classes successfully clear a set of filters that assure
15
270
1K
3
1
15
Amara's Law:
We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.
"In the short-term things change less than we think. In the long-term, things will change more than we think."
9
62
590
0
2
14
.
@swyx
: "are we running out of data⁉️"
.
@ce_zhang
: "on earth? there is still more data. some of it isn't openly available"
Food for thought:
Will AI models knowledge saturate soon~ish?
Then what?
Regardless, great
@latentspacepod
with
@togethercompute
:
1
7
15
Six LLMs-related tech advancements that will usher in the era of AI:
1> LLMs information grounding and referencing (Bing's chatbot is a primal example)
2> Efficiently connecting LLMs to tools (such as databases, simulators, calculators and various apis)
4/N
1
1
15
@svpino
This trend will grow during 24'.
Eventually 'data' + 'natural language' would be used to create a big portion of the overall real-world software
0
0
15