Today we're excited to announce that we're partnering with
@CarperAI
of Stability on bringing the first RLHF-trained GPT-3 like model to the open source community.
This will be huge. Let us explain
Today we're excited to announce public access to Humanloop for Large Language Models!
We're making it easier than ever to build incredible products with GPT-3
Sign-up at
TogetherAI have brought the context window of LLaMA-2-7B to 32k tokens.
This is potentially the first time an open source model with competitive performance to ChatGPT can be fine-tuned for long-context tasks.
This is important for the following reasons 🧵
The playground may be the first way you interact with GPT-3... but it's also an IDE in disguise.
We want to transform it to be best way to build useful apps with LLMs. Here's how ↓
There are hundreds of different use cases for Large Language Models like GPT-3.
Starting today we're going to be showcasing a different awesome start-up and what they're building every day.
RLHF – Reinforcement Learning from Human Preferences.
Models are fine tuned using RL from human feedback. They become more helpful, less harmful and they show a huge leap in performance. An RLHF model was preferred over a 100x larger base GPT-3 model.
Our CEO
@RazRazcle
sat down with
@rowghani
of YC to discuss:
• the future of Large Language Models
• the huge opportunities for startups
• and how to build differentiated products in the AI space
Video link below ↓
Cohere has found that data pruning on a pre-training corpus for LLMs, can result in a performance improvement while using just 30% of the training data.
Here’s why this is important 🧵
OpenAI is the most important company in the world right now.
We sat down with
@sama
where he revealed their future plans.
He shared the ambitious vision for ChatGPT, announced plans for a new stateful API and revealed what's held back multimodal GPT.
The 100k context window Claude is now available in the Humanloop Playground.
Try it out, paste in a book... we're excited to see what you build with it!
Model pricing details from Anthropic below
⚡️ Streaming the response from GPT-3 a huge UX win
Feels magical as you watch the AI respond.
Now available on the Humanloop Playground and Generate endpoints so you can use it in your app.
Unveiling the new !
Humanloop offers you the best way to rapidly build and continuously improve your NLP models. Powered by Active Learning.
The best ML teams follow this approach. Here's how you can too
CarperAI will be building in public, and will be releasing the data, code and weights over the coming months. Follow along in their discord.
Like what happened to Stable Diffusion, we can't wait to see the innovation that happens once these models become available for all
These models show far greater ability to take instruction, which has massively increased their usability.
We think RLHF-tuned models will ultimately be applied to every domain and task, and these systems will unlock incredible amounts of value in the real world.
For more prompting guidance have a look at - we hope the tips and tricks help you on your journey of building effective LLM applications.
Thank you to
@prof_oz
for the guest post.
OpenAI have just released text-davinci-003
• higher quality writing
• handles more complex instructions
• better long form content generation
and now supported in
@Humanloop
playground!
What's the difference between an AI engineer and an ML engineer?
"The AI engineer looks after the 0-1 phase whereas the ML Engineer takes on the 1-N phase"
@swyx
penned the viral essay 'Rise of the AI Engineer' () and is the founder of the
@aiDotEngineer
Big News: We've just added Evaluation Functions to Humanloop!
This gives you a powerful new set of tools to evaluate model performance, allowing you to build LLM applications with much greater confidence than before.
Here’s why this is a big deal 🧵
This week we locked ourselves in a country home in South England.
Our goal: Launch a unique AI-powered product.
• 4 teams
• 48 hours
• 1 rule - everything had to be built on Humanloop
Here's the results 🧵
New Release: We’ve just added OpenAI Function Calling to Humanloop!
Function Calling has been hailed as one of the most significant OpenAI feature releases of 2023. However it has been tricky to get them into production.
Here's how Humanloop helps solve this ⬇️
You can now talk to GPT-4 in the Humanloop discord!
The OpenAI live demo inspired us, so we used GPT-4 to create a GPT-4 bot!
With the help of GPT-4 it only took
@jordnb
about 20 minutes to code this from scratch!
We're excited to host
@ycombinator
's London meet-up at
@localglobevc
's office this summer for future founders to network and learn about startups. We'll be telling the story of starting Humanloop. If you're in London, sign up through
@startupschool
at .
All OpenAI models are down right now. Follow here for updates
(Anthropic, Cohere, AI21 are all running well. An excellent time to toggle on a backup deployment in Humanloop)
Want to understand how Humanloop can help you build differentiated products with LLMs?
Join a live demo with
@RazRazcle
↓
Today, 6pm GMT / 10am PT
March 7th, 5pm GMT / 9am PT
March 9th, 6pm GMT / 10am PT
LLMs are the next big computing platform. At least as important as the internet, perhaps more. We want to empower the next million developers to build AI-first applications
If you want to build successful products and apps with GPT-3, sign up now.
Like any good
@ycombinator
company, we
@humanloop
are always trying to listen to our customers and build what they want.
Here are 5 recent feature releases driven directly by customer requests and feedback:
1) Multi-provider chat-focused playground
Human feedback is critical to aligning these models to do as you want.
We've built the specialised tools needed for this crucial component and are delighted to be contributing to this open-source effort.
One of the most valuable talks from at
@OpenAI
DevDay was given by
@john__allard
and
@colintjarvis
. Yet many seemed to have missed it.
In this blog,
@jordnb
breaks down its key insights to show how you can start Maximizing the Performance of LLMs.
Building an LLM app has evolved significantly since 2022.
If you're working on an AI project, you should familiarize yourself with the anatomy of a modern LLM application. Here's a quick overview:
• LLM model - the core reasoning engine; an API into
@OpenAI
,
@AnthropicAI
,
2. Few shot learning
For extra clarity, provide examples.
This is particularly useful if you want an uncommon tone, style or syntax. If accuracy is paramount, stuff the whole context window with examples.
Fine tuned models at scale
Instead of a 4-5x higher per-token fee, you pay an hourly rate for hosting fine tuned models.
This gets economical at scale. And at scale you'll definitely want to be fine tuning.
(though wow at the jump for hosting Davinci vs Curie)
🚨 Announcement 🚨
Just a few hours ago
@AnthropicAI
dropped Claude 3 - a new suite of models which outperform GPT-4 and Gemini Ultra.
We're excited to announce that Claude 3 is now available on Humanloop. Bring your API Key to test, evaluate and deploy
Big product update: We’ve introduced support for multi-modal models!
You can now input images and text to models like GPT-4V in our Editor or API.
This presents a new frontier in the range of apps that can now be built with AI.
Here’s what GPT-4V looks like in our Editor ⬇️
Here’s how you connect an LLM to tools to give it extra capabilities.
For example, include google search in your context window with simple syntax like {{ google(query) }}.
Read more here:
Will Larger Context Windows Kill Retrieval Augmented Generation?
As model providers continue to extend their context windows, researchers have been looking into whether or not this enhances performance over Retrieval Augmented Generation (RAG).
The HELM benchmark from Stanford is a fantastic resource to understand the relative strengths of large language models
Here’s the overall ranking on the 16 core tests with the win-rate over the other models.
Announcing Holistic Evaluation of Language Models (HELM) v0.2.0 with updated results on the new
@OpenAI
,
@AI21Labs
, and
@CohereAI
models. HELM now evaluates 34 prominent language models in a standardized way on 42 scenarios x 7 metrics.
Subtle update you'll have missed:
Finetuned gpt-3.5-turbo is now only 3x of the cost of the base 3.5-turbo. It was 8x before.
Finetuning just got even more compelling.
Our feature with
@EO__Global
is now live on YouTube 🚀
Watch
@jordnb
and
@RazRazcle
talk about their journey to co-founding Humanloop, building an MVP at
@ycombinator
and their advice for AI startups
In
@siliconrepublic
today!
Our CTO Peter Hayes chats to
@VishGain
about Humanloop's London origins, getting into
@ycombinator
expanding to San Francisco and how we help enterprises confidently build with AI 🚀
3. Plan for your instructions to be taken very literally
These models have been trained to predict the most likely sequences of words. Protect around simplistic pattern matching, and be careful what you wish for.
Happy Tuesday: Today we're launching support for
@Meta
's Llama 2 in the Humanloop Playground!
You can now go to and start testing out the model right away! No need for an API Key or any other configuration.
Here's what this means ⬇️
but going from a cool demo () to a fully-fledged production app is still too hard.
• hard to evaluate
• hard to make reliable (hallucinations, repetition)
• hard to get real domain knowledge into these models
New version of Programmatic dropping today! 🎉
You can now add ground truth labels directly in the UI! 🤯
Programmatic uses these to give you instantaneous feedback on precision and recall.
👉 Check it out:
SO excited to share what we've been working on!
We're building the best tool for programmatic labelling.
It's a super powerful way to build your datasets for NLP and it's getting great feedback from early users.
Sign up for access 👉
In fact there is a new job title emerging for generalist engineers who have a strong familiarity with LLMs and AI tooling: "The AI Engineer"
@swyx
Despite the claims that 'AI is coming for the jobs of software engineers first’, so far it has only created a greater need for them.
The Allen Institute for AI (
@allen_ai
) just released a 3T token dataset to support open training of large language models.
This is the largest
#opensource
dataset released this year, with 50% more tokens than Llama2’s training data.
Here’s what you need to know 🧵
Love machine learning?
Join a YC-backed startup that bridges the gap between people and machine learning, allowing people to truly understand and control their models.
➡
#machinelearning
#activelearning
#python
#PyTorch
Retweets appreciated! 😃
Thanks
@nasscom
or including Humanloop in their Generative AI Landscape Map!
We're building out tooling to support the next generation of applications that will be built on top of large language models.
It's great to be recognised here alongside so many amazing companies in AI!
Humanloop takes the guesswork out of prompt engineering and model development with
• Feedback at scale
• Experiment tracking
• One click fine-tuning
all brought together with a simple SDK.
1. Feedback
Eyeballing a few examples isn't enough for production applications. So, Humanloop lets you collect end-user feedback at scale, unlocking actionable insights on how to improve your models.
Discover the issues that you are missing!
Want
#machinelearning
in
#Zapier
?
Humanloop can get to the same level of accuracy as traditional supervised learning with ✨10% ✨ of the labelling requirements.
➡
Can
@AnthropicAI
's Claude 2 model beat OpenAI's GPT-4?!
It has:
- Access to data up to 2023
- 100,000 token context window
- Improved performance in coding, math and reasoning
We've added Claude 2 to Humanloop's Playground so you can compare them side-by-side and see for
3. Finetune
Prompts can only take you so far.
Finetune GPT-3 with your data to get a custom model at lower latency and lower cost.
With Humanloop, you can do it with a single click and no code.
(we spun this up at a hackathon last week, using Humanloop as the quickest way to go from idea → prompt template → deployed GPT-3).
Install the plugin here to supercharge Obsidian with AI
We're excited to be working with
@CarperAI
, the biggest group in the open-source RLHF space! Their recently released TRLX library makes it easy to adapt LLMs from human feedback. Check it out!
One consistent piece of advice that has comes on almost every episode of the High Agency Podcast...
• Start building today❗
@sourcegraph
is one of the world's most popular AI coding tools. Their CTO and Cofounder
@beyang
shares his thoughts on this ⬇️
An excellent summary of the history and recent progress in NLP by
@chrmanning
:
Strikingly he dates the present era of NLP as starting in 2018! We are just at the start of what's possible
WHAT THEY DO
Phind is a next-generation search engine.
They don't just return a list of links, they directly actually answer your question in natural language. With citations.
AI tool of the week: YC Idea Matcher
It's an open-source semantic search engine for matching ideas with
@ycombinator
companies!
Just give it an idea and it links you to similar YC backed startups.
5. Battle test it with data
To really trust it, you'll need to see how it works with diverse data. Set up some scripts to iterate through data points or use a tool like
@humanloop
to help you evaluate it on real world data.
Exciting update to Copilot on Perplexity with
@OpenAI
's GPT-3.5 fine-tuning API! We've improved speed, reduced cost, and matched GPT-4 performance. Copilot now engages faster without losing quality. Here's what you need to know: 🧵
If you're building with large language models and want to go from:
Impressive demo -> Impressive Product
Sign up to get early access to our closed beta.