🌊 Software eating Software

Apr 16, 2023

👋 Hello! I’m Ivan. Join >5.7K entrepreneurs surfing startup waves.

Summary

Hello startup riders! This week’s good stuff for your startup brain includes:

🧮 Software eating software: developer superpowers.
🍬 Startup Candy: Steve Jobs on managers vs team leads.
💵 Deals: 28 startup deals in Spain (>€250M).

✍️ Software eating Software

People are excited about large language models, I’m bullish on the “AI giving developers superpowers” angle. I believe its the biggest lever in the short-run.

Even though developer tools have made code production more efficient over the years, the software assembly line is still pretty primitive.

I believe the way the sausage is made is about to change pretty radically.

😟 Problem

For programmers - Programming languages are still a very rudimentary way of giving instructions to machines.

For users - “Traditional” Graphical User Interfaces (GUIs) are still a very rudimentary way of communicating with machines.

🤩 Solution

Text is the universal interface. It can encode all human knowledge. We’ll use it.

For programmers - giving instructions to machines on our own terms, using Machine Copilots:

For users - interacting with machines on our own terms using Conversational User Interfaces (CUIs)

⏳ How we got here

The history of Machine Learning in a nutshell:

1930s: Alan Turing, inspired by biology, comes up with the idea of building a computer modelling a human brain.
1955: Clark and Farley from MiT built a structure of neural units using computers.
1956: Dartmouth Conference, considered the birth of AI as a field of study. Early AI research focused on rule-based systems and symbolic reasoning.
1959: Oliver Selfridge proposed a model for how machines could process uncategorized information like humans do. Scary title: “Pandemonium: A Paradigm for Learning.”
1980s: AI experiences a period of hype and excitement, known as the "AI winter," followed by a decline in funding and interest in AI research.
2010s: In 2015 ImageNet, a repository of millions of carefully curated and labeled images, is released by Fei-Fei Li and Andrej Karpathy - with the firs deep learning models successfully identifying images. Yet, while progress in computer vision was taking off, developments in natural language processing were still lagging.
2017: Google researchers publish “Attention is all you need” and change the field radically. The transformer model in a nutshell:
- The transformer architecture designs a more efficient communication protocol between neurons that allows key decisions to be made faster.
- Instead of breaking an input into smaller chunks, and processing them sequentially, the transformer is structured in such a way that every element in the input data can connect to every other element.
- Each layer decides which inputs to “pay attention to” while analyzing text.
2020s: The new Software assembly line emerges. Developers gain superpowers. Ethical implications of AI become a growing concern, including bias, privacy, and job displacement.

🚀 Market Map

Copilots

Interesting enabling startups

Codeball: AI powered code review.
Grit: Fix technical debt automatically.
Stenography: Document entire codebases. Every time you hit save.
GradientJ: Build NLP Applications in Minutes
Adrenaline: Stop using Stack Overflow for help. Talk directly to your codebase like you would an expert.
Waveline: Build and integrate AI in minutes, no ML experience required.
Baseplate: backend for LLM apps.
Second: Connect a developer bot to an existing project, or create a new one, and see a significant boost in product development.
Vellum: Bring LLM-powered features to production with tools for prompt engineering, version control, back-testing, performance monitoring, and continuous fine-tuning. Compatible across all major LLM providers.
Helicone: Track costs, usage, and latency for GPT applications with one line of code.
Wild Moose: Solve production incidents in minutes instead of hours. Delegate the laborious search through logs, changes, metrics & code to lightning-fast AI.
Berri: API for SaaS businesses to create ChatGPT apps programmatically
Buildt: Engineers at companies like Stripe and Airbnb have to work with million-line codebases; our LLM-powered tool makes this simple
Stack: Enable any developer to build with Large Language Models through fine-tuning and composability.
Rubbrhand: ML Training in 1 Line of Code

If you are a spanish founder / based in Spain building in this space I’d love a chat.

✔ Conclusions

10x Production: Software production speed is going to 10x.
From GUIs to CUIs: Graphical User Interfaces will be replaced by Conversational User Interfaces. Opens an opportunity to disrupt big incumbents.
More devs, not less: We are about to witness an avalanche of new age “Developers” as the barriers to entry change / adapt to LLMs.
More AI Alignment: There’s a lot of heat on both sides of the debate. This is the best thing I’ve read on the subject, published in 2015. I also like this PoV by Gary Marcus (ex Uber AI Labs, MiT), seems balanced and timely:

I am not worried, immediately, about “AGI risk” (the risk of superintelligent machines beyond our control), in the near term I am worried about what I will call “MAI risk”—Mediocre AI that is unreliable (a la Bing and GPT-4) but widely deployed—both in terms of the sheer number of people using it, and in terms of the access that the software has to the world.

🐇 Follow the White Rabbit 🕳️

The AI Revolution: The Road to Superintelligence - Tim Urban
The AI Revolution: Our Immortality or Extinction - Tim Urban
The Age of AI - Bill Gates
Software 2.0. - Andrej Karpathy
A Short history of Artificial Intelligence
AI risk ≠ AGI risk - Gary Marcus
Text is the universal interface - Roon
Developer tools 2.0. - Sequoia
The industrial revolution of software
Neural Networks
Generative AI Exploding
Alternatives to Github co-pilot

💭 Thinking

“One of the painful things about our time is that those who feel certainty are stupid, and those with any imagination and understanding are filled with doubt and indecision.”
—Bertrand Russel