AI Learnings - January 5-7, 2026
AI Learnings - January 5-7, 2026
Overview
A focused day covering Claude Code & Workflows and AI Agents & Orchestration.
Claude Code & Workflows
- @krispuckett: "The FOMO around Claude Code is real"
- @tonylongname: "@clawdbot is now my homeโs PM:"
- @davekiss: "clawdbot update: rebuilt my entire personal site via telegram last night while watching netflix in bed"
- @tobi_bsf: "Using @clawdbot for a week now and it genuinely feels like early AGI"
- @ctjlewis: "Everyone who tried Claude Code over the holidays and has been thinking deeply about this for 7 days:"
AI Agents & Orchestration
- @rauchg: "The primary lesson from the actually successful agents so far is the return to Unix fundamentals: file systems, shells, processes & CLIs"
- @Hesamation: "Amazon proved small language models can outperform 500x bigger LLMs in agentic tool calling"
Other Highlights
- @OpusClip: "๐ Turn podcasts into viral clips in minutes"
- @lamxnt: "How to stop feeling behind with AI (practical guide)"
- @mattpocockuk: "If you feel like you're caring less about code quality in the AI age, read this"
- @DanielleFong: "listen, if you get all your notes into the AI it dramatically enhances its own ability to 'get' what you want because it has associative memory and access to that"
- @pamelafox: "I always get asked how I keep up with gen AI news, so I've updated my blog post about 'how I learn gen AI' with my fav bloggers"
Key Takeaways
1. Claude Code continues to reshape how developers approach coding
2. Agent orchestration patterns are maturing with new tools and frameworks
---
Curated from 14 postsSource Posts
M
If you feel like you're caring less about code quality in the AI age, read this
โ
that's it... Amazon proved small language models can outperform 500x bigger LLMs in agentic tool calling. NVIDIA already showed this idea before, and this is another proof that the right fine-tuning can make SLMs better in agentic applications than LLMs, with a fraction of the cost.
agent-focused companies must adopt more the development (dataset, evals, post-training, design) of LLMs this year. they have the data and the right playgrounds for these LLMs. it's economically senseless to use proprietary large models in most agentic use cases.
read this paper and also NVIDIA's ("Small Language Models are the Future of Agentic AI") to be convinced.
O
๐ Turn podcasts into viral clips in minutes. Virality Score finds your best moments, AI B-Roll adds flair, Speaker Detection keeps faces framed, and removes filler words like โumโ & โuh.โ
Don't sleep on this. Try OpusClip today and get 25 free clips/month.
d
@clawdbot is now my homeโs PM:
Clawd is in discord with wife and me.
โ Either of us drops topics anytime ("add to the list: kitchen reno")
โ Clawd researches and refines with that person
โ They accumulate in a shared file
โ Saturday 10pm: Clawd prepares brief, refreshes research, prioritizes items.
โ Sunday 9am: Clawd sends roundup to our Discord
โ We discuss, mark done, archive
thanks @steipete !!
G
The primary lesson from the actually successful agents so far is the return to Unix fundamentals: file systems, shells, processes & CLIs.
Don't fight the models, embrace the abstractions they're tuned for. Bash is all you need.
P
I always get asked how I keep up with gen AI news, so I've updated my blog post about "how I learn gen AI" with my fav bloggers.
https://t.co/3oElwllkr7
You can also follow them here: @simonw, @GergelyOrosz, @HamelHusain, @intellectronica, @isaac_flath, @GaryMarcus https://t.co/02etDiN3zT
D
clawdbot update: rebuilt my entire personal site via telegram last night while watching netflix in bed
notion โ astro, 18 posts migrated, dns moved to cloudflare. never opened my laptop.
https://t.co/3K6U8kHNUO
N
To move from a weekend AI demo to an AI production-grade application, you need to architect these 4 layers ๐๏ธ๐๏ธ https://t.co/EzQDEqt8kZ
K
The FOMO around Claude Code is real.
You don't need a project idea. You need to describe your work and let it interview you.
So many people are posting rad Claude Code wins: apps, sites, workflows. But most people are staring at a terminal thinking "I don't even know what's possible."
Try this:
1. Open Claude Code
2. Dump everything about your workโyour work/role, tools you touch daily, tasks you repeat, stuff that annoys you, wild ideas you've always wanted to try, your passions, your hobbies, etc
3. Paste this:
"Based on what I shared, ask me 5-7 questions to understand my workflow better. Then suggest 3 things I could build, ranked by impact vs complexity. Use the AskUserQuestion tool to help me"
Answer honestly (specifics = better suggestions)
4. Pick the one that makes you go "wait, that's possible?"
The skill isn't coding. It's describing friction, your wild ideas, asking what it can learn about you to help you.
Claude Code isn't automatically a magic wand. You wouldn't hand someone a hammer and say "build something." You'd explain your situation and ask what's possible.
Always happy to jam/chat about this stuff.
A
Andy Orsow
@andyorsow
Feeling like I should be using Claude Code but have no idea exactly what I should be using it for. Just a bundle of non-technical FOMO over here.
D
listen, if you get all your notes into the AI it dramatically enhances its own ability to "get" what you want because it has associative memory and access to that. it is a massive capabilities unlock to have this. for the things i have been dreaming about i can reference them by gesturing at it.
and the problem of handling a mountain of context is inherent to the process, because it eventually creates it. it's perfectly reasonable people are bootstrapping and dogfooding here
T
Using @clawdbot for a week now and it genuinely feels like early AGI.
The gap between "what I can imagine" and "what actually works" has never been smaller.
But: Token consumption is insane. Running a personal assistant 24/7 hits limits fast.
Two things need to happen:
โข Agents get smarter about token usage (just shipped a PR reducing prompt overhead)
โข Model prices keep dropping (already happening - GPT/Claude/Gemini racing down)
The future where everyone has a personal AI running in the background isn't limited by capability anymore. It's limited by cost.
That window is closing fast. ๐
r
How to stop feeling behind with AI (practical guide)
M
$150,000/Week with Zero Employees: How He Absolutely Cooked With AI
L
Everyone who tried Claude Code over the holidays and has been thinking deeply about this for 7 days:
All white collar jobs are toast. It is over