Software 3.0 and the AI Engineer Landscape (talk notes & slides)

category: note slug: ai-landscape

2024 update

I presented an updated version of this talk at CascadiaJS 2024! slides here, video pending.

2023 talk

I was honored to be invited back to Infobip Shift for the 3rd year in a row, and this time to my total surprise and honor I was given the keynote stage right after John Romero:

It's still unreal to see my slides and "artwork" on a big stage like that, and for people in Europe to fully get it even if they dont live this at work yet. The Shift twitter account person was particularly good at notetaking live. I was also randomly interviewed for Al Jazeera by a wonderfully friendly reporter.

Abstract

Everyone has gone AI mad, and there is a lot of noise, hype, and demos, but not a lot of guidance on practical usecases. We will go over a mental model of who is doing what and explain both where the opportunities are and where the non-AI developer can start going down the rabbit hole.

Slides

https://docs.google.com/presentation/d/1WPULCJsbdsYLTEnAzkdVtqKRDdx4MrfMH04u5HAD4Wc/edit?usp=sharing

Presented a shorter form version of it at Daily.dev

20 min version presented at React Summit US

Video

{% youtube https://www.youtube.com/watch?v=q68lG7ubHvc %}

and 20 minute React Summit version

{% youtube https://www.youtube.com/watch?v=NaKCw2p382s %}

Raw notes

Target 25 mins

5 mins: DATA Timeline, data proof points
- madness
  - AI AI I/O
  - agents rise
  - no adoption in enterprise? https://twitter.com/bentossell/status/1686387000260358145
    - HN quotes
  - sarah guo quote
- People
- publication speed overwhelming
- AGI foom https://x.com/swyx/status/1653887955474460672?s=20
  - and AGI fears -> wait but why, not kill everyone memes
  - https://x.com/swyx/status/1653885657448841216?s=20
- open source AI model development https://x.com/swyx/status/1653891330668658689?s=20 and transformers variants https://x.com/swyx/status/1653890389039013889?s=20
- chegg textbooks and stackoverflow collapsing
- Who's making money from AI
  - Jasper
  - Midjourney
  - Copilot
  - ChatGPT
  - Character.ai
  - Agents?
  - Case Studies
    - Levelsio
    - Headlime
    - Zapier AI
    - Notion AI
    - Vercel AI
5 mins: TREND: Rise of the AI Engineer
- What is Software 3.0?
  - Software 1.0 vs 2.0
  - Software 3.0 -> code is impt
  - ???
- https://huyenchip.com/2023/08/16/llm-research-open-challenges.html 1. Reduce and measure hallucinations
  2. Optimize context length and context construction
  3. Incorporate other data modalities
  4. Make LLMs faster and cheaper
  5. Design a new model architecture
  6. Develop GPU alternatives
  7. Make agents usable
  8. Improve learning from human preference
  9. Improve the efficiency of the chat interface
  10. Build LLMs for non-English languages
- Disciplines of the AI Engineer
  - Prompt Engineering & Structured Responses
  - AI UX
  - OSS AI - Hosting & Infra
    - beware local LLM u curve
  - OSS AI - Finetuning & Evals
  - AI Productivity Devtools
  - Tooling for AI
  - AI Agents
5 mins: EXAMPLES: Latent Space Highlights
- Fundamentals: Datasets, Benchmarks, Algorithms
- Celebs: Geohot, Chris Lattner, Code Interpreter, Rise of AI Engineer, Replit, Reverse Prompt Eng
- Unintuitive lessons
  - Bitter Lesson - encoding things is a temporary boost
  - RAG - second brain
  - Code always better
    - Replit model
  - Larger not always better
    - Chinchilla - Llama
    - Phi-1 data paper, tinystories
    - we have tricks to scale - Geohot 1.7tr
    - tshirt sizing of models
  - Raw datasets are overrated
    - dont need to keep updating every year
    - Deep learning on datasets -> 4 kinds of training
    - https://twitter.com/altryne/status/1661236778458832896
    - llama 2 - estimated cost 2.5b
  - Agents are the future
5 mins: Landscape presentation
- equal weighting categories is mistaken
  - https://twitter.com/swyx/status/1701472499840532853/photo/2
- survey
- Foundation Model
  - openai
  - anthropic
- Libraries/Frameworks
  - LangChain
  - LlamaIndex
  - Deepset Haystack
  - Vercel AI
  - Semantic Kernel
  - Guardrails
  - Guidance
- Vector DBs
  - Vector DBs
- OSS Hosting/Training
  - ?
- Misc
  - Prompt management
  - Monitoring/observability
- 5 mins Latent Space University
  - reuse stuff from cityjs talk
  - 0: learn how to prompt
    1. OpenAI GPT3 API Basics (completed)
    2. what is a token https://twitter.com/swyx/status/1653893171787747329
    3. Prompt Tooling and Memory (completed)
    4. Code Generation with GPT3 (completed)
    5. Image Generation with Dall-E, Stability AI, Lexica, and Midjourney (in production)
    6. Speech-to-Text with Whisper (tbd)
    7. Finetuning & Running Open Source Models (tbd)
    8. AI Agents
YOU ARE NOT TOO OLD
- https://www.latent.space/p/not-old
- Opportunities of the AI Engineer
  - Prompt Engineering
  - AI UX
  - OSS AI - Hosting & Infra
  - OSS AI - Finetuning & Evals
  - AI Productivity Devtools
  - Tooling for AI
  - AI Agents
- consider
  - Ethics, Sourcing
  - security - software 2.0 security
  - Multimodality
- What to rule out:
  - AGI hard
  - Google vs OpenAI vs Meta wars
  - Xrisk, Regulation, Law
  - Threadbois

Context

I was honored to be a keynote for this year's Infobip Shift conference, going right after the legendary John Romero:

swyxio / swyxdotio