Lewpen retweeted
Google's engineers stopped letting AI grade its own homework, and they open-sourced how
it's called the Elephant-Goldfish Model, and Google says thousands of its engineers run it every day on production code
the elephant is one long-running session that remembers everything: the architecture, the intent, every small decision.
it never writes code. it interviews you, argues with you and keeps the record
the goldfish are fresh sessions with zero memory. they never saw the design talk, so they don't inherit its blind spots
the workflow:
1. brainstorm - you and the elephant debate options under a strict no-code rule
2. draft - a markdown design doc with the problem, the plan, the alternatives and the exact files to change
3. explain back - one goldfish has to explain the design. if it gets it wrong, the doc is ambiguous and goes back
4. attack - several goldfish hunt edge cases, hidden dependencies and logic gaps
5. go/no-go - a last goldfish checks that every file path, data structure and function signature is spelled out
AI made coding faster and made slop faster too. this moves the attack to before the first line of code exists
the trade-offs: it only runs inside Antigravity for now, and the default reviewer list skips security, so you add your own security goldfish
free, Apache 2.0. the model writes the code, the harness decides what ships, and how one engineer actually works with agents day to day is below:
Finally summarized what I've been sending to friends in DMs View article
Lewpen retweeted
DeepSeek open-sourced dsh-superagent, and it works through a repo's GitHub issues and pull requests with many agents at once
the goal is written into the design: take as little human attention as possible and deliver as much value as possible
why that matters: a person can follow 3 to 5 active sessions at once. one task is capped by the model, multitasking is capped by your attention
how it's built:
> a long-lived router talks to humans and never runs tasks itself
> isolated workers run the tasks
> events like a new issue or a mention trigger the work, so nobody types the next prompt
> a watchdog brings the whole thing back after a crash
the part worth stealing: the router works without a bash tool, and a lighter model usually handles it fine
their own instance reads public group chats, and even when nobody tags it, it finds the root cause and opens the pull request
the full write-up behind the design is below:
Lewpen retweeted
DeepSeek Harness went from v0.1 to 1.54 million daily users in 83 days
roughly 92 trillion tokens a day, counting users on the API
the part that matters for harness engineering: about 35,000 third-party plugins exist, and about 60% of active users have built or installed at least one
the design behind it fits in one line: everything is a plugin, and built-in features can be replaced by third-party ones
the team's report with all the numbers is below:
Lewpen retweeted
Harvard now teaches AI agents all the way down to the GPU
the course is CS2680, Modern AI Systems, and it starts where most tutorials stop
the top of the stack comes first:
> Claude Code and how it runs
> the agent loop
> evals
> what every token costs
then it goes down one layer at a time:
> model deployment
> routing
> KV cache
> quantization
> speculative decoding
the clever part is the homework. all 5 assignments take apart one agent system: first you study how it works, then you build its harness yourself, then you optimize it down to the inference layer
it answers the question most AI courses skip: how does a modern agent get from the app you type into to the hardware it runs on
if you want a job in agent infra, this is the syllabus to copy, and the harness is where every assignment starts
the loops and graphs that sit on top of that stack are below ↓
Lewpen retweeted
Same prompt. Two top AI models.
One made a poster. The other made a movie.
Gemini 4 Argon vs Opus 5.5. Here’s my honest take:
GEMINI 4 ARGON
• Strong first impression
• Bold, stylized, poster-like
• Plays it safe with composition
OPUS 5.5
• Slower to impress, then wins on detail
• Better sense of weight, light and scale
• Takes more creative risks
Gemini gives you the thumbnail.
Opus gives you the scene.
Quick social clip? Argon.
Something that should feel like a film? Opus.
Models don’t just differ in quality. They differ in taste.
Whose taste do you trust more?
Lewpen retweeted
Prompt timing in AI video is still broken. 🧵👇
We fed the exact same 10-second multi-action prompt to three flagship models: Gemini Omni 1.2 Flash, Seedance 2.5, and Kling 4.0.
The goal? A black cat inspecting an ornament (0-3s), knocking it off to bounce on the rug (3-6s), and climbing into the Christmas tree to knock it over (6-10s).
Here is what actually happened:
🔹 Gemini Omni 1.2 Flash: Nailed the static scene composition, but struggled with prompt sequence adhereance. The timing logic collapsed mid-sequence, keeping the cat grounded rather than initiating the final dynamic crash.
🔹 Seedance 2.5: Massive jump in physics fidelity. It caught the micro-interactions perfectly — the paw tapping the ornament mid-air feels fluid and natural without typical limb distortion.
🔹 Kling 4.0: Best temporal awareness out of the three. It correctly executed the precise time-stamps (0-3s target lock ➡️ 3-6s paw interaction), retaining heavy volumetric lighting and skeletal weight throughout the move.
The takeaway: Models are finally solving organic animal movement, but chronological instruction following remains the ultimate filter separating cinematic tools from random generators. Seedance and Kling are currently taking the lead in action-reaction loops.
Check out the side-by-side comparison video above 👆
Exact prompt provided in the first reply ⬇️
Lewpen retweeted
Static AI video is officially dead. 💀
We spent a month testing flagship models (Seedance & Kling) under complex physics stress-tests. Here is why multi-action prompts are the only way to survive retention drop-off: 👇
Lewpen retweeted
Gemini 4 Argon vs Kimi K3.1 Pro 🦚🔥
Grounded physical accuracy vs cinematic volumetric glow.
I broke down the exact frame-by-frame physics and trade-offs in the full benchmark article below ⬇️
Which pipeline fits your workflow? 👇
Lewpen retweeted
🚀 Introducing RoboMoCap!
We are releasing the largest locomotion-centric egocentric vision × optical motion capture dataset for humanoid robot learning, joint effort with @GroundedSI @BitRobotNetwork @NokovMocap
📊200+ hours of recordings
🏃 290+ diverse actions
👓 Egocentric Vision
🎥 120 Hz optical motion capture with ~10 mm tracking accuracy
🌐 bitrobot-foundation.github.i…
🧵👇
Lewpen retweeted
Had my AI dig through millions of startup job postings to compile these startup hiring statistics for 2026:
📈 Fastest growing roles
👉 Forward Deployed Engineer +218% YoY
👉 Creative Designer +55.6%
👉 Video Producer +50.0%
📉 Fastest declining roles
👉 PHP Developer -74.1%
👉 Content Editor -56.7%
👉 Graphic Designer -56.0%
👩💻Remote work
👉 Down from its pandemic 37.9% peak to 16.9%
👉 Hybrid at 27.9% of roles
💰 Salaries in US
👉 Best-paid role: Engineering Manager with a median salary of $228k a year
👉 Best-paid city for Software Engineers: Santa Clara ($225k)
🗺️ Where startups are hiring
🇺🇸 United States is the top country by far
🌁 San Francisco is the top city, closely followed by London, UK
More data with charts here:
startup.jobs/startup-hiring-…
Lewpen retweeted
Google is now literally showing ecommerce businesses specifically how to get traffic from AI Search.
Yes, directly inside your Google dashboard.
This is probably one of the most useful AI Search updates for retailers in 2026.
Let’s go through it.
If you want to see whether your business is appearing across Google AI, ChatGPT, Claude, Perplexity, Grok and broader AI search, start here. It’s free:
seo-stuff.com/free-audit
Google recently made its new AI Performance Insights report generally available inside Merchant Center for eligible businesses in the United States, Canada, Australia, India and New Zealand.
The current report covers English-language conversational shopping queries across Google AI Mode and AI Overviews.
Google is essentially giving ecommerce companies an AI Search diagnostic report.
You can see whether your products are being surfaced when somebody uses Google’s AI features to research what to buy, and where competitors are getting visibility that you are not.
Let’s start with share of voice.
Google calculates AI share of voice based on how many AI impressions your brand or products receive compared with you and the competitor set Google has defined for related searches.
What exactly does that mean?
Well, when Google AI is helping shoppers research your category, how much of that visibility are YOU getting?
If competitors are appearing constantly while your products barely show up, you finally have first-party Google data showing you the gap.
No manually asking AI Mode the same question 50 times or trying to infer everything from referral traffic - Google itself is telling you.
The report then breaks conversational shopping searches into three stages:
Discovery.
Evaluation.
Ready to buy.
This might be the most useful part of the entire report.
Suppose you have strong visibility during Discovery, but it falls during Evaluation and nearly disappears at Ready to buy.
That tells you your products are being surfaced early in the research journey, but losing visibility as shoppers get closer to making a decision.
The opposite could happen too.
Maybe you perform well for shoppers who already know what they want, but competitors are being introduced much earlier.
That is far more useful than simply knowing you were “mentioned in AI Search.”
And it is why SEO Stuff approaches AI Search Optimization as more than checking whether ChatGPT mentioned your name once:
seo-stuff.com/gold-plan-pack…
Then there are the actual searches.
Traditional ecommerce SEO has historically focused heavily on relatively clean keywords like “running shoes,” “carry-on luggage” or “standing desk.”
AI shopping queries can look more like:
“What’s a good lightweight running shoe under $150 for someone with flat feet who runs mostly on pavement?”
Those searches contain use cases, constraints, features, price ranges and preferences all at once.
Google’s report is built around that reality.
Its Top Terms section shows the concepts shoppers are prioritizing inside conversational queries.
Google gives examples such as “maximum cushioning” and “arch support,” then tells merchants to incorporate relevant high-frequency terms into product titles and descriptions where appropriate.
That is Google directly connecting conversational AI Search behavior to product optimization.
The report also surfaces Popular Attributes.
Think size, color, material and other structured product specifications.
Google can show which attributes shoppers frequently care about and which may be missing from your product data.
If shoppers repeatedly want to know whether something is machine washable, expandable, 100% cotton or supports a certain weight, but your product data does not clearly provide that information, you have an information gap.
Google’s recommendation is straightforward: populate missing attributes and prioritize the ones shoppers are asking about most.
Then there is Top Search Intents.
This goes beyond individual terms and looks at WHY somebody is searching.
A shopper may want an office chair for lower back pain, want to compare two models, stay under $300 or understand whether mesh or leather is better.
Those are very different buying situations, and Google is now giving retailers more information about which conversational shopping intents are actually occurring.
So the optimization loop becomes pretty straightforward:
Look at your AI Search share of voice, identify where competitors are beating you, see where visibility falls across Discovery, Evaluation and Ready to buy, review the terms and attributes shoppers care about, improve your product information and keep monitoring.
Google is explicitly recommending this kind of workflow.
Keep product data accurate.
Add relevant high-frequency terms where appropriate.
Populate missing attributes.
Keep reviewing the report as new opportunities appear.
There is another important detail.
The AI Performance report currently measures organic AI visibility.
Paid Ads traffic is not included.
That makes this particularly relevant for SEO.
But Merchant Center product data is only part of the equation.
People researching expensive or complicated products also ask things like:
“What is the best X for Y?”
“X vs. Y.”
“Is X worth the money?”
“Best X under $500.”
“What should I look for when buying X?”
A product feed alone may not answer those questions, but your website can.
Buying guides, comparison pages, category pages, FAQs, use-case content, product education, original research and expert explanations give Google more context around the buying decision.
Suppose shoppers are researching “best running shoes for plantar fasciitis.”
Your product feed can help Google understand which shoes may be relevant.
Your website can explain which features matter, which products contain those features, how models compare and who each shoe is designed for.
That gives Google considerably more information to work with across the broader research journey.
This is what the SEO Stuff Premium Content Bundle is designed to build:
seo-stuff.com/premium-conten…
Then there is authority.
Your product feed and website are still information coming from YOU.
Search and AI systems also encounter reviews, editorial publications, comparison articles, industry websites, third-party mentions and contextual backlinks discussing your company and products.
That broader authority is why the SEO Stuff done-for-you package combines content with contextual DR50+ placements:
seo-stuff.com/gold-plan-pack…
And Google just gave ecommerce businesses considerably more data to see where those efforts may need to be focused.
There is another development worth paying attention to - Google is also introducing Conversational Attributes in Merchant Center.
These are optional product-data fields designed to help AI systems and conversational agents understand more nuanced information about products.
They can include product questions and answers, supporting documents, related products, variant information and popularity information.
Google recently disclosed a test involving lululemon.
The company submitted conversational product attributes, and Google says those attributes were incorporated into relevant AI Mode product recommendations 50% of the time during testing.
Mind you, that does NOT mean adding conversational attributes gives you a 50% chance of being recommended, it just means the additional information lululemon supplied was incorporated into relevant recommendations half the time within that test.
Google wants more detailed product information because AI shopping queries are becoming more detailed, and it is increasingly giving merchants tools to both supply that information and measure what happens afterward.
If you run an ecommerce company, I would be paying very close attention to this report.
AI shopping is already measurable inside Google, and Google is now giving businesses tools specifically designed to help them understand and improve their presence within it.
And if you want to see where your business stands across Google AI, ChatGPT, Claude, Perplexity, Grok and broader AI search, start here:
seo-stuff.com/free-audit
You can win a customer on Google and still lose them when they ask ChatGPT whether they should buy.
In fact, nearly 60% of customers now say that talking to ChatGPT, Claude, Gemini and so forth convinced them *not* to purchase something.
That means there is a huge opportunity opening up for businesses right now.
Let’s go through it.
And if you want to see what ChatGPT, Claude, Google AI, Perplexity, Grok and broader AI search are currently saying about your business, start here. It’s free:
seo-stuff.com/free-audit
Recently Semrush’s Exploding Topics team recently surveyed 2,338 U.S. consumers about how they use AI during the buying process.
One number really stood out.
57.5% of AI users said information from an AI chatbot had caused them to decide not to buy something.
Among people who use AI at least weekly, that jumped to 69.7%.
Basically, you can rank #1 on Google, run the ad, get the click, build the landing page, publish the case studies and convince the customer that your product might solve their problem.
Then the customer opens ChatGPT, Gemini, Claude or Perplexity for a second opinion - and potentially changes their mind.
That is a completely new part of the conversion funnel that most businesses are barely measuring.
The influence works in both directions too.
57.51% of AI users said they had purchased something following an AI chatbot recommendation, while 59.27% said an AI recommendation had introduced them to a brand or product they did not previously know about.
That creates a pretty enormous opportunity, obviously.
Imagine somebody is about to buy from Company A.
Before paying, they ask:
“Is Company A the best option for a 20-person marketing agency?”
The AI researches the category, finds comparisons, reviews, third-party articles, pricing information and use cases.
And suddenly:
“You may also want to consider Company B.”
Company B just entered a buying decision it was not even part of five minutes earlier.
That may increasingly be one of the most valuable parts of AI Search Optimization.
And that is what SEO Stuff helps businsses with every single day.
seo-stuff.com/gold-plan-pack…
There is another number businesses should pay attention to.
47.54% of all consumers surveyed said they at least sometimes consult an AI chatbot for information about a company or seller before making a purchase.
Among AI users, that rises to 65.42%.
So the AI “background check” is becoming a real part of the buying journey.
Someone hears about your company and asks whether the business is legitimate, what the reviews are like, what the downsides are, who the competitors are and whether there is a better option.
Your potential customer may be researching you in an environment you do not control.
Your broader digital footprint matters more than ever.
Depending on the platform and query, AI systems may draw on your website, reviews, comparison pages, publications, directories, forums, news coverage and other available sources.
So there are really two sides to the problem.
First, your own website needs to clearly explain what you do, who you serve, what things cost, how you compare, what use cases you support and what results customers have achieved.
Give AI systems accurate information to work with.
But your website is not the entire internet.
The second part is what exists about you elsewhere: credible editorial coverage, relevant backlinks, reviews, third-party mentions, independent comparisons and industry references.
Those sources give AI systems additional evidence about who you are and where you fit.
That is exactly why SEO Stuff combines content and authority:
seo-stuff.com/gold-plan-pack…
The content builds out the questions buyers are asking, while the authority side strengthens the third-party footprint around the brand.
And this becomes especially important for non-branded questions.
There is a huge difference between:
“Is Salesforce good?”
and:
“What is the best CRM for a 15-person B2B company that needs HubSpot integration and does not want to spend more than $500 per month?”
The second question creates room for brands the buyer may never have considered.
Businesses should be answering questions like best X for Y, X vs. Y, alternatives to X, how much does X cost, is X worth it, who is X best for, what are the disadvantages and what should I look for before choosing?
There is also an important Google angle.
65.01% of AI users surveyed said they had replaced at least some product-related Google searches with AI chatbot conversations, but the study did not find that people had simply abandoned Google.
For many consumers, AI is complementing it.
That is why I do not think businesses should choose between SEO and AI Search Optimization.
You need both.
And if you want to see where your business stands across ChatGPT, Claude, Google AI, Perplexity, Grok and broader AI search right now, start here:
seo-stuff.com/free-audit
Lewpen retweeted
Introducing patch: build and play games.
- Every game dev tool in one CLI and MCP
- 50,000+ open-source assets
- Remix any game on patch, or mod the ones on your PC
- Ship it to everyone on patch, opensource :)
We're opening it to everyone today: patch.gg
Lewpen retweeted
MiniMax-H3 Facial Realism
CloseUp realism LoRA. Captures subtle textures and micro-expressions
- trained on high-quality human-centric portrait
- sharp, photorealistic close-ups
huggingface.co/prithivMLmods…
Lewpen retweeted
introducing the world's first fully ai video streaming service!
built using minimax h3 max turbo, you can now watch entire movies and easily create your own ones
Lewpen retweeted
Dreamina Seedance 2.5 is unbelievable!
We're burning through an ungodly amount of credits for our feature film, Nexus.
But there's a trick to saving money with 30-second outputs that you need to know.
Let me show you the secret steps and prompts👇🏼🧵
Lewpen retweeted
You can now run a 2.8 TRILLION parameter model on a 4GB GPU.
Someone open-sourced a tool that uses "Layer-wise Inference." It only loads one layer onto your GPU at a time. so the VRAM you need depends on the layer size, not the model size.
no quantization. no distillation. no pruning.
→ DeepSeek-V3 (671B) on 12GB
→ Llama 3.1 405B on 8GB
→ Kimi K3 (2.8 TRILLION params) on under 4GB
→ works with almost every open model
the biggest model on it needs the LEAST VRAM. K3 is sparse MoE, so it streams only the experts a token actually routes to instead of a whole dense layer.
2.8 trillion parameters running in less VRAM than a 70B.
Lewpen retweeted
we're putting together a directory of hardware gadgets that you can vibe code your own firmware for
→ buildinpublic.com/gadgets
what are YOU building?
Lewpen retweeted
Se acabó... nadie puede decir ahora que esto es IA.
Te enseño a crear vídeos como este ⬇️
🤖 Made with AI
Lewpen retweeted
🚨 Major studios spent tens of millions of dollars on live-action versions of classic cartoons
A 22-year-old animator from China spotted this trend exploding on Douyin and made $13,942 last month by being the first to bring it to Western TikTok and Reels.
Turning old hand-drawn cartoons into realistic footage started over there and is still climbing. He copied the format and launched it before anyone else.
> Kimi K3: hold the series in one window, write 250 shot prompts off a single character file: 1 day
> Seedream: draw every still with the build copied word for word, never paraphrased: 20 min
> Seedance: animate the keepers only, so a wrong face never reaches a render bill: 40 min
> ElevenLabs: one locked voice, a delivery note under every line: auto
Studios need soundstages, real animals and a full crew for one scene.
He just needs a $67 stack and his own cat.
Every prompt is in the article above 👇
🤖 Made with AI



































