Creative & Visual Media
Ilker (@ailker) demonstrated a hyper-realistic video of pyramids created in one day using generative AI models on FAL, predicting its use in future SuperBowl commercials.
'pyramids are real. this video isn't'
I made this in just one day using generative AI models on @fal. It’s not perfect, but I can already see how the 2027 #SuperBowl commercials will be made. pic.twitter.com/FNDkLygUH3
— ilker (@ailker) February 7, 2026
In the past few weeks I’ve been cold emailed by a 15-year-old and 16-year-old offering significant technical insights and contributions.
We will see within our lifetime the rise of young supergeniuses that far surpass previous generations in both intellect and demonstrable…
— Guillermo Rauch (@rauchg) February 7, 2026
Javi Lopez (@javilopen) highlighted advancements in cinematic video generation for video games, suggesting the next step is full films.
Cinematic scenes for video games = SOLVED ✅
Next steps = FILMS https://t.co/C3IMFdsfrR
— Javi Lopez ⛩️ (@javilopen) February 7, 2026
I always felt CFG was a patch to fix a training problem we didn't yet understand.
Training with only normal distributed noise teaches the model that each step will have a perfectly normalized error from the previous step, which is not the case. Therefore, it is incapable of… pic.twitter.com/9Wn3d60PbI
— Ostris (@ostrisai) February 7, 2026
Ostris (@ostrisai) shared insights on improving diffusion models by fine-tuning a LoRA on Z-Image without CFG, resulting in higher quality images that avoid over-saturation.
I always felt CFG was a patch to fix a training problem we didn't yet understand.
Training with only normal distributed noise teaches the model that each step will have a perfectly normalized error from the previous step, which is not the case. Therefore, it is incapable of… pic.twitter.com/9Wn3d60PbI
— Ostris (@ostrisai) February 7, 2026
Alex Volkov (@altryne) noted Bytedance’s upcoming Severance 2.0 as a major leap in video generation.
Bytedance is about to release severance 2.0 and the early test show it's going to be another step function in video gen
— Alex Volkov (Thursd/AI) (@altryne) February 7, 2026
Heather Cooper (@HBCoop_) teased Kling 3.0 capabilities with a video demo.
Trigger finger ready…
– Kling 3.0 pic.twitter.com/FYB3yDAmJ6
— Heather Cooper (@HBCoop_) February 7, 2026
fofr (@fofrAI) showcased Waymo’s generative World Model for hyper-realistic autonomous driving simulations using Google DeepMind’s Genie 3.
https://t.co/njw9npLzMt pic.twitter.com/oukxdx45Xu
— fofr (@fofrAI) February 7, 2026
https://t.co/AwOrcadkK7 pic.twitter.com/SoMPHiRoQN
— fofr (@fofrAI) February 7, 2026
A.I.Warper highlighted the emerging trend of multishot techniques in AI video generation, sharing a demo video and expressing excitement to experiment further.
Excited to try this one.
Seems multishot is the current trend for Q1
— A.I.Warper (@AIWarper) February 7, 2026
Generative AI video models are advancing rapidly, enabling realistic content like a one-day pyramid video indistinguishable from reality using tools on FAL, predicting SuperBowl-level commercials soon.
'pyramids are real. this video isn't'
I made this in just one day using generative AI models on @fal. It’s not perfect, but I can already see how the 2027 #SuperBowl commercials will be made. pic.twitter.com/FNDkLygUH3
— ilker (@ailker) February 7, 2026
AI-generated ads are proliferating on TikTok, comprising about 1/3 of feeds, featuring hyper-realistic spokespeople driving purchases as viewers inquire about product sizes in comments.
It’s happening.
I’d guess that ~1/3 of the ads I get on TikTok now are AI generated.
And they’re working – the comments are all people asking what size to get! pic.twitter.com/VQmuodNmEW
— Justine Moore (@venturetwins) February 7, 2026
With video models now so capable, differentiation requires compelling storytelling rather than tricks, returning focus to “idea guys.”
We're back in the age of the idea guy.
AI video models have gotten so good that you can't impress people with cheap tricks or pure novelty anymore.
You have to craft a real story that viewers want to follow…which is hard! https://t.co/RtWX1JqoAj
— Justine Moore (@venturetwins) February 7, 2026
Software Development
Omar Sar praised Anthropic’s 2.5x faster version of Claude Opus 4.6, available via Claude Code and the API, noting no compromises in quality and its value for rapid iteration, debugging, and production applications.
2.5x faster and without any compromise on quality and capabilities
i wasn't complaining about speed with opus 4.6, but this feels extremely useful for rapid iteration and tasks like debugging, and when you need to kill bugs fast for your prod applications https://t.co/XPkK7IOmi7 pic.twitter.com/um5HGSfOJf
— elvis (@omarsar0) February 7, 2026
A radical vision for fully automated software factories eliminates human intervention, reimagining development processes with AI.
A genuinely radical approach to software development with AI, without any human intervention. Even if this approach doesn’t work for many cases, I think we need more leapfrogging visions for how to redo processes with AI: https://t.co/GjkJ31wGOA
See also: https://t.co/2rh7a1MLkG pic.twitter.com/mesPEesYc1
— Ethan Mollick (@emollick) February 7, 2026
Builders are fine-tuning specialized “orchestrator” models using Claude Code, focusing on clear reasoning and tool selection for tasks like documentation or image generation.
I'm using Claude Code to fine-tune an uncensored open source model.
I am training an "Orchestrator Model." This model acts like a human would do and only does 2 things.
1. Think clearly without overcomplicating the thoughts. (reasoning)
2. Find best tools to execute the… pic.twitter.com/2AxHK0ElMm
— CJ Zafir (@cjzafir) February 7, 2026
Reliance on AI coding tools like Claude Code, Cursor, and Codex has surged, with true builders unable to complete projects without them.
If the answer is YES, you’re not going hard enough. https://t.co/5cxLt9H0eQ
— Riley Brown (@rileybrown) February 7, 2026
Automation & Orchestration
Guillermo Rauch (@rauchg) discussed how AI agents empower individuals with the productivity of large teams, accelerating human potential.
In the past few weeks I’ve been cold emailed by a 15-year-old and 16-year-old offering significant technical insights and contributions.
We will see within our lifetime the rise of young supergeniuses that far surpass previous generations in both intellect and demonstrable…
— Guillermo Rauch (@rauchg) February 7, 2026
Emerging orchestrator models bridge chatbots and IDEs by reasoning simply, selecting optimal tools (e.g., Claude Code for building, Gemini for images), and self-improving via datasets and memory graphs for scalable, personalized execution—addressing gaps in tools like OpenClaw.
I'm using Claude Code to fine-tune an uncensored open source model.
I am training an "Orchestrator Model." This model acts like a human would do and only does 2 things.
1. Think clearly without overcomplicating the thoughts. (reasoning)
2. Find best tools to execute the… pic.twitter.com/2AxHK0ElMm
— CJ Zafir (@cjzafir) February 7, 2026
Strategy & Ecosystem
Rauch also predicted the rise of young “supergeniuses” enabled by AI self-teaching tools across disciplines.
In the past few weeks I’ve been cold emailed by a 15-year-old and 16-year-old offering significant technical insights and contributions.
We will see within our lifetime the rise of young supergeniuses that far surpass previous generations in both intellect and demonstrable…
— Guillermo Rauch (@rauchg) February 7, 2026
Volkov speculated on Anthropic’s Claude Opus 4.6 release strategy, including a faster version for Claude Code and API, potentially rebranded from Sonnet 5.
> be Anthropic
> train sonnet 5
> mfw it's better than opus and also faster> send it to companies to announce release
> moneybags.gif
> rename to Opus 4.6
> slow it down
> undercut OAI 5.3 release> wait
> release original Sonnet speed with an up-charge> profit $$$ https://t.co/8xHJnrMhMr
— Alex Volkov (Thursd/AI) (@altryne) February 7, 2026
lol fist they give us Opus 4.6 and now they speed it up?
Hmmmm
Was this sonnet 5 all along 🤔 https://t.co/8xHJnrLJWT
— Alex Volkov (Thursd/AI) (@altryne) February 7, 2026
Cristóbal Valenzuela (@c_valenzuelab) likened European AI-by-fax tech to broader challenges in integrating AI with legacy systems.
This is a metaphor for many of those AI integrations with legacy software out there https://t.co/osm0aeP6Kx
— Cristóbal Valenzuela (@c_valenzuelab) February 7, 2026
Dan Shipper observed significant drops in SaaS stocks following recent AI model launches, attributing it to short-term panic similar to reactions after DeepSeek’s release. He argued that software is set for fundamental changes but most major SaaS firms will endure, linking to his 2023 Every article “What Comes After SaaS.”
saas stocks are down big after the model launches this week
almost 3 years ago i wrote an article for @every called "what comes after saas": https://t.co/YbjsO67rYH
i think software is going to fundamentally going to change over the next few years and i think this panic is…
— Dan Shipper 📧 (@danshipper) February 7, 2026
Kimi K2.5 has become the most popular model on OpenRouter.
Kimi K2.5 is now the most popular model on OpenRouter pic.twitter.com/kftInjEt8u
— OpenRouter (@OpenRouterAI) February 7, 2026
Mac-only releases for AI tools are a growing annoyance, alienating gamers and corporate users alike.
Mac only releases for AI tools is an annoying trend. This is an opinion that unites both gamers and corporate users.
— Ethan Mollick (@emollick) February 7, 2026
In the AI era, audience and distribution are exponentially more valuable as moats.
Audience & distribution are becoming exponentially more valuable by the day.
— The Boring Marketer (@boringmarketer) February 7, 2026
Building AI companies mirrors high-frequency trading, rapidly exploiting fleeting arbitrage opportunities.
building companies with ai is like being a high frequency trader (and will become moreso) observing machines attacking small windows of arbitrage until they close, and moving to the next bet
— The Boring Marketer (@boringmarketer) February 7, 2026
Advanced role definitions in prompts act as a “cheat code” for unique AI outputs by activating niche expertise, outperforming generic engineering personas.
here's how i get AI outputs that nobody else gets…
i play with role definition and it's a CHEAT code
"you're a software engineer" activates the most generic, averaged-out training data possible
thousands of people prompted the exact same thing before you… you're basically…
— Machina (@EXM7777) February 7, 2026
Many X articles exhibit hallmarks of low-quality AI generation, resembling conspiracy theories or productivity hacks.
Why do so many X articles read like AI-written conspiracy theories/health hacks/LinkedIn productivity tricks that would fall apart with any serious research?
On second thought, this isn’t really a question.
— Ethan Mollick (@emollick) February 7, 2026
Skills trained on specialized assets, like copywriting websites, can 10x output quality.
you can just 10x your AI copywriting by building a skill trained on assets from this website… pic.twitter.com/u3nmsOynHZ
— Machina (@EXM7777) February 7, 2026