Vibe Coding and AI Coding Agents
Simon Willison shared details on building a “vibe-coded” web browser project released by Cursor, noting it now compiles after fixes and sharing screenshots from his Mac build.
In case anyone's interested, that vibe-coded(ish) web browser project that Cursor released the other day does actually compile now, they fixed it up and added instructions to the README. I built it on my Mac and took these screenshots: pic.twitter.com/YTDtW2r4tB
— Simon Willison (@simonw) January 18, 2026
More in this gist https://t.co/SxpnFLfg7Z
— Simon Willison (@simonw) January 18, 2026
Levelsio highlighted a YouTuber running multiple Claude Code terminals in a “cluster” to vibe code apps aiming for $1M revenue, calling it the most interesting shipping effort recently.
This guy is running a cluster of Claude Code terminals vibe coding apps until he hits $1,000,000
Most interesting person shipping I've seen recently
He's on here too @matthewmillerai but doesn't seem to tweet a lothttps://t.co/2K3973Ngv1 https://t.co/iJjLPWplac pic.twitter.com/pOUnuetSRA
— @levelsio (@levelsio) January 18, 2026
Dan Shipper emphasized it’s the best time to be an “ideas guy” due to AI tools enabling rapid prototyping without deep coding skills.
there has never been a better time in history to be an ideas guy
— Dan Shipper 📧 (@danshipper) January 18, 2026
AI Video Generation and Character Manipulation
Levelsio demonstrated Mocap on PhotoAI.com, integrating models like Nano Banana Pro, Wan 2.2 Animate, and ElevenLabs for cheap AI video creation (10 credits/video at a loss leader).
Here's how Mocap works on my app https://t.co/1vEawpI5vb
It integrates Nano Banana Pro + Wan 2.2 Animate + the ElevenLabs voice library all in one feature
I run all of those models at loss now, so it's probably one of the cheapest place to make AI content now, 10 credits per… pic.twitter.com/zB1WJP3cNC
— @levelsio (@levelsio) January 18, 2026
He showcased his “e-girl vlog” using character swaps to explain the tech, preserving motion while altering appearance.
My first vlog as an e-girl where I explain how character swaps work https://t.co/WhkfYuzc60 pic.twitter.com/rZXplJaVbu
— @levelsio (@levelsio) January 18, 2026
Cocktailpeanut detailed lipsyncing videos with LTX-2, preserving originals and editing only mouths via masking in WanGP.
Lipsync Any Video using LTX-2
1. The original video remains untouched
2. Only the lips are modifiedIf I understand correctly, unlike the usual lipsync AI that generates entirely new videos from images, this preserves your original video while editing just the mouth movements https://t.co/7N5vd6gZrv
— cocktail peanut (@cocktailpeanut) January 18, 2026
Prompt Engineering as a Skill for Video Models
Cocktailpeanut stressed detailed prompts are key with LTX-2’s Gemma text encoder, shifting video gen success from model limits to user skill (“skill issue”).
Geat idea. I've also had this issue time to time.
Because LTX-2 uses a good text encoder (Google's Gemma), it's really good at understanding nuanced, detailed prompts. Detailed prompts matter.
At this point, generating a good video is not a model issue. It's a skill issue! https://t.co/j4xmz0ZA8e pic.twitter.com/xiXpwZICBD
— cocktail peanut (@cocktailpeanut) January 18, 2026
He equated advanced text prompts to LoRAs for concept extraction, chainable without fine-tuning.
Most people think "text prompt" means text-to-video, text-to-image, etc.
But the real power (which I'm talking about here) is text-to-concept.
Text prompts NOT used for output generation, but used for on-the-fly concept extraction, chainable. https://t.co/1CZFX0hcCd
— cocktail peanut (@cocktailpeanut) January 18, 2026
LLM Agent Memory Mechanisms
Omar Sar highlighted a survey on evolving memory systems for LLM agents, covering effective implementations.
A Survey on the Evolution of LLM Agent Memory Mechanisms
Nice survey on some of the effective memory mechanisms used with LLM-based agents. pic.twitter.com/hv66zD8wb6
— elvis (@omarsar0) January 18, 2026
Economic Strategies Ahead of AGI
Levelsio observed friends “speedrunning” wealth accumulation via assets (stocks, real estate, gold) to be capital-heavy before AGI disrupts labor markets.
So everyone around me is speedrunning making as much money as fast as possible and spending it on buying assets (stocks, ETFs, commodities like gold and silver, real estate etc) to be asset heavy when the AGI hits https://t.co/lRFpnqFAwE
— @levelsio (@levelsio) January 18, 2026
Vibe Coding and AI-Driven Software Development
Discussions highlighted “vibe coding” as a transformative skill, turning consumers into engineers by leveraging intuitive AI prompts for building apps and tools.
"Vibe coding turns every consumer into an engineer" pic.twitter.com/YgGMCJ0XuD
— Justine Moore (@venturetwins) January 19, 2026
My overall opinion on vibe coding has changed in the past year. I used to think there would be a a niche of vibe coders, similar to indie hackers. But I’m not sure anymore. I think within one year, vibe coding will just be a mandatory skill in business.
— Riley Brown (@rileybrown) January 17, 2026
Ethan Mollick showcased daily AI-generated game demos using Claude Code, such as a spaceship puzzle game and SocFight where sociologists battle with management theory mechanics.
Still commanding AI to make one weird game demo a day.
This one: "A game where you need to stop a spaceship from crashing into the sun and your only tool is future Claude Code." The entire design & all puzzles 100% AI, I gave minor UX feedback.
Play: https://t.co/0syGFxQX1V pic.twitter.com/dmGCHLPni2
— Ethan Mollick (@emollick) January 19, 2026
Continuing to build a game a day by just asking the AI.
This is SocFight: a fighting game where historical sociologists beat each other up. Will Max Weber's Iron Cage attack defeat Durkheim's dangerous Anomie?
(Asked Claude Code to use OpenAI's image generator when needed). pic.twitter.com/V4nOMiInGy
— Ethan Mollick (@emollick) January 18, 2026
Continuing to have AI build a weird game demo a day. Here is: "Make a game where you have to prevent the apocalypse, but the interface is just Jira tickets"
Pretty fun/funny branching storyline, all text is AI created with minor feedback from me. Play: https://t.co/Zr5OM7z3FN pic.twitter.com/wkQhX2zIo8
— Ethan Mollick (@emollick) January 17, 2026
Riley Brown emphasized vibe coding’s shift to a business essential, planning internal tools to replace SaaS platforms and seeking YouTubers for education.
BTW if you’re a YouTuber passionate about vibe coding and want to help educate millions of people on how to vibecode apps and (soon) agents.l, My DMs are open. Bonus if you are technical and are good at explaining complex topics simply. Don’t care as much about followers and… https://t.co/Csu47BHcss
— Riley Brown (@rileybrown) January 19, 2026
We will get 150M long form video views this year without spending a dollar on influencers. (Or more) Educational high value content, that compounds over time.
In house, educational content machine.
Whenever we think “should we pay influencers?” we think… is there a way to…
— Riley Brown (@rileybrown) January 19, 2026
Tomorrow is day one of my "Vibe Coding for Business" series.
As a founder in charge of growth i'm subscribed to over 20 different scattered software platforms.
I'm convinced I can replace 90-95% of these apps with my own internal tools that I vibecode.
Then, at the end of…
— Riley Brown (@rileybrown) January 17, 2026
Machina shared a GitHub with 40+ Claude Code tips and promoted laziness in rebuilding systems via AI patterns.
this github has 40+ Claude Code tips from beginner to expert: pic.twitter.com/EEI8k3zzCG
— Machina (@EXM7777) January 19, 2026
AI rewards laziness more than hard work and nobody wants to admit it…
think about what you can do right now:
– reverse engineer any system you see online
– rebuild entire business models in hours
– skip months of trial and error
– copy successful frameworks and adapt them…— Machina (@EXM7777) January 18, 2026
Advances in Generative Media and LoRAs
ilker praised Flux.2 [klein] 4B for training high-quality LoRAs for outpainting, zooming, and object removal, noting surprising consistency from a smaller model.
👀 trained 4+ LoRAs with flux klein 4B today and, the consistency & quality surprised me. didn't expect results this good from a smaller model.
gonna share how i generated the datasets tomorrow but trust me it's easier than you think https://t.co/dczCwYzrs4
— ilker (@ailker) January 19, 2026
Justine Moore pointed to AI video trends like epoxy floor renovations and teeth rendering as remaining challenges, calling teeth the “final frontier.”
Epoxy floors are the latest AI renovation trend – and they’re extremely satisfying to watch pic.twitter.com/j1ehOmFuVw
— Justine Moore (@venturetwins) January 19, 2026
Teeth are the final frontier for AI video.
Models give everyone these perfect white chiclet teeth – and they often fuzz out when in motion.
This is a great example of the phenomenon. It’s one of the last “tells” that a video was generated. pic.twitter.com/sYWhYrCoZ4
— Justine Moore (@venturetwins) January 18, 2026
The “Wait Equation” and AI Laziness Strategies
Mollick revisited his “Wait Equation,” advising delays on projects like software until AI agents improve, citing rapid progress in coding and video.
This turned out to be true. Many software projects should have waited until agentic coding tools got to today’s level
Given trajectory in coding, agents, video, etc. there are lots of other projects that you should be lazy about today since it is faster to wait for AI to improve https://t.co/SopfegLFx3
— Ethan Mollick (@emollick) January 19, 2026
Machina echoed this, arguing AI favors laziness through pattern recognition and prompting over grinding from scratch.
AI rewards laziness more than hard work and nobody wants to admit it…
think about what you can do right now:
– reverse engineer any system you see online
– rebuild entire business models in hours
– skip months of trial and error
– copy successful frameworks and adapt them…— Machina (@EXM7777) January 18, 2026
The Boring Marketer shared AI prompts for finding unique product angles to boost sales.
the difference between "why isn't this selling" and "we can't keep up with demand" is usually the right angle
how to find your unique angle (paste this into your AI etc) pic.twitter.com/vYHPJS7RsI
— The Boring Marketer (@boringmarketer) January 19, 2026
AI Model Breakthroughs in Math and Agents
Mollick noted GPT-5.2 Pro solving multiple Erdos problems, a threshold breach from o1’s era, often with human prompting and Lean iteration.
Erdos problems are a definite example of models breaching a threshold. The idea that an AI could solve one, let alone many, on its own would have been insane a year ago (o1 was brand new), and now we have multiple Erdos problems solved by GPT-5.2 Pro in the last couple weeks. https://t.co/hTjSVVm0eL
— Ethan Mollick (@emollick) January 18, 2026
To be clear: GPT-5.2 Pro is not solving these autonomously, it is prompted by a person, and it often iterating using Lean.
— Ethan Mollick (@emollick) January 18, 2026
He critiqued cute agent interfaces as dead ends, favoring project management tools, and called for serious voice modes to manage agents.
The fact that all of the big AI voice modes are powered by dumb models, let alone sycophantic dumb models that are designed to have disfluencies that fake a human chat (“um”), undersells the value of voice in managing agents.
A “serious voice mode” for work would be very useful
— Ethan Mollick (@emollick) January 18, 2026
As someone who made one of those cute gamey agent interfaces that went viral, I do think it is likely that they are a dead end.
Even if tasks are not fully automated, as roon suggests, we already have tools for delegating long-running tasks. It will look like project management. https://t.co/l861N50n84
— Ethan Mollick (@emollick) January 18, 2026
OpenRouterAI promoted agent skills for optimizing calls across harnesses.
`npx add-skill OpenRouterTeam/agent-skills`
Instantly optimize your OpenRouter calls.
Supports 8+ agent harnesses pic.twitter.com/qhEYAcPIpj
— OpenRouter (@OpenRouterAI) January 17, 2026
AI in Business, Marketing, and Content
Machina advocated recording client calls for AI analysis via NotebookLM to refine pitches and content, and shared TikTok agency scaling to $37k MRR with AI slop.
you're sitting on your most valuable data source and not using it…
if you run an agency, every call with a lead or customer should be recorded
here's what i do:
> record every single interaction
> dump it into NotebookLM
> let it analyze patterns in objections, questions,…— Machina (@EXM7777) January 18, 2026
just hired 3 interns to upload AI slop on 150 tiktok accounts… we're sitting at $37k MRR so far pic.twitter.com/4M3LDrteyB
— Machina (@EXM7777) January 17, 2026
Riley Brown pushed in-house educational video for 150M views, bypassing influencers.
We will get 150M long form video views this year without spending a dollar on influencers. (Or more) Educational high value content, that compounds over time.
In house, educational content machine.
Whenever we think “should we pay influencers?” we think… is there a way to…
— Riley Brown (@rileybrown) January 19, 2026
Machina stressed sales skills over other degrees in an AI era.
if you're gonna waste 5 years at uni… study sales
everything else is obsolete
but sales teaches you stuff you can't get from online courses:
– reading people in real time
– handling objections on the spot
– calibrating your art based on micro-reactionsthe education system…
— Machina (@EXM7777) January 18, 2026