AI/LLM Links of the Week

Trying something new -- you may be familiar with my "If you only read one thing" weekly posts, but if for some reason you're hungry for more, here it is!

I read way too much every week, and picking just one thing to highlight is brutal (there's so much going on), but those daily newsletters are a total firehose. Let's see if this splits the difference.

AI/LLM Links of the Week?

If you have to read just one: The Future of AI is Vertical by Euclid Ventures https://insights.euclid.vc/p/the-future-of-ai-is-vertical

Why I clicked:

• Foundation models are inherently general and beat specialized models at most tasks, so this is a bit contrarian (but I like it)

What I learned:

• Generally, vertical specific data and workflows can be defensible — training a vertical-specific model likely isn’t.
• Software dev is one of the most obvious verticals to chase - there are lots of specialized workflows and tools already, and compilers and linters make correctness relatively attainable, compared to other domains where you’d need astronomical amounts of training data to get to confidence.

On to the rest:

Navigating the generative AI disruption in software by McKinsey Tech, Media, Telecom practice https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insig[…]igating-the-generative-ai-disruption-in-software?cid=soc-web

Why I clicked: It’s McKinsey

What I learned: They estimate $35-40 billion in new custom software spend, because AI can help you build whatever you might have bought and customized in the Before Times. Several other excellent graphs

Building AI Agents: Lessons learned over the past year by Patrick Dougherty https://medium.com/@cpdough/building-ai-agents-lessons-learned-over-the-past-year-41dc4725d8e5

Why I clicked: Agents!

What I learned: Your agent config is not a moat — but things like agent-model interfaces, data connectors, UX, memory and evals might be

Announcing: Tools

https://x.com/AnthropicAI/status/1796210547077128578 https://huggingface.co/spaces/huggingchat/chat-ui/discussions/470

Why it matters: Anthropic and HuggingFace both announced new tool interfaces — Huggingface’s is provocative!? It only works for one model now (Cohere Command-R+) but the intent is to generalize…

AI will make money sooner than you’d think by Nilay Patel, Decoder https://www.theverge.com/24173858/ai-cohere-aidan-gomez-money-revenue-llm-transformers-enterprise-stochastic-parrot

Why I clicked: Cohere is a big fish you don’t hear from much

What I learned: Enterprise adoption is happening, it’s slow but it’s real

What Apple’s AI tells us: Experimental Models by Ethan Mollick https://www.oneusefulthing.org/p/what-apples-ai-tells-us-experimental

What I learned: There does seem to be a meaningful opportunity for small models with excellent context — we are seeing a similar effect with Stride Conductor. Most of Conductor's secret sauce is about setting the best possible context for LLMs to write code, and doesn’t ask as much of the model itself.

The Latent.Space Podcast with Mike Conover of Brightwave https://www.latent.space/p/brightwave

What I learned: The limits on output (4096 tokens) are more relevant than limits on input window for most LLMs — this can be worked around by breaking down tasks as small as possible.? This sounds familiar! (We’ve run into this on recent engagements)

The rise of medium code by Nick Schrock https://dagster.io/blog/the-rise-of-medium-code

Why I clicked: I love the term — not low-code, but medium-code — low code tools are limiting, custom code and DevOps are hard, maybe medium code is the best of both worlds?

What I learned: I like the idea of coarse-grained containers -- allow practitioners to focus on business logic, but use existing code toolchains to inspect and harden and ship

Introducing Apple’s Foundational Models by Apple ML Research https://machinelearning.apple.com/research/introducing-apple-foundation-models

Why I clicked: I’d like to know how my iPhone is about to work

What I learned: Apple is using techniques like LoRA fine-tunes and quantization that came out of the open source hobbyist community — it explains how they can get it all so small! (But I guess it's still time for a 15 Pro)

Smart Paste for context-aware adjustments to pasted code by Google Research https://research.google/blog/smart-paste-for-context-aware-adjustments-to-pasted-code/

What I'm thinking: This is a fresh take on Copilot for copy/paste — Google saw that this behavior was happening, used LLMs to respond to it, turned it on for their devs and have seen steady adoption — how could we do this for other common workflows in our own teams?

Programming is mostly thinking by Tim Ottinger https://agileotter.blogspot.com/2014/09/programming-is-mostly-thinking.html

Why I clicked: Title checks out

What I’m thinking: How does this change workflows and the role of developers, if the coding and the thinking are both AI-assisted?

Generative AI is not going to build your engineering team for you by Charity Majors https://stackoverflow.blog/2024/06/10/generative-ai-is-not-going-to-build-your-engineering-team-for-you/

Why I clicked: Clickbait title + StackOverflow = popcorn emoji

What I’m thinking: It is hard to build great teams, and AI makes it harder in the short term — how do we make it easier?

The ARC prize for AGI progress https://arcprize.org/blog/launch

Why I clicked: $1M prize!

What I learned: This is a great way to highlight some observed limitations of LLMs — and a clear benchmark to know if we actually overcome them

Just for fun: AI will become mathematicians’ Co-Pilot by Scientific American https://www.scientificamerican.com/article/ai-will-become-mathematicians-co-pilot/

Why I clicked: Terence Tao is a big deal!

What I learned: T minus 3 years to “Copilot for theoretical math”