makerzz Trends · Wednesday, 2 September 2026
Tiny Models Beat Big Ones, and Creators Are Making Content About It
Small transformers outperform LLMs on ARC, agents edit your video, and fantasy football drafts flood every feed
September opens with a weird flex: hobbyist-scale models embarrassing billion-parameter giants, Google Play picking fights with open-source apps, and fantasy football search volume swallowing every other trend. Here is what creators are actually posting about it, and what you should make today.
What creators are doing
YouTube556 HN points"I trained a tiny model that beats GPT" explainer format. A solo dev published a blog post about training a small transformer in 1.5 hours that outperforms many LLMs on the ARC benchmark — and the Hacker News thread hit 556 points with 148 comments. Creators in the AI-education space are turning this into "David vs Goliath" walkthrough videos: screen-record the training, show the benchmark scores side by side, narrate the surprise.
Why it works: Underdog stories with receipts get shares. The concrete claim (small model > big model) is inherently clickable and the 1.5-hour training time makes it feel reproducible.
Copy this
Screen-record yourself replicating the training on your own laptop, show the benchmark comparison, and title it 'My $0 model just beat [big model name]'.
Make this today
The 'Tiny Model Beats Big Model' Reel
- 1
Pick your David-vs-Goliath angle10 min
Use the ARC benchmark story (small transformer trained in 1.5 hours beats many LLMs) as your reference. Write a hook: 'I spent 90 minutes training a model that outperforms [big name].' You do not need to actually train the model — you can narrate over the original blog screenshots.
- 2
Grab 4-5 screenshots from the source5 min
Go to mvakde.github.io/blog/44-on-arc-1/ and screenshot the training setup, the benchmark comparison table, and the results. These become your B-roll or carousel slides.
- 3
Record a 60-second voiceover or talking head15 min
Structure: hook (5 sec), context on what ARC is (10 sec), the training process (15 sec), the results with specific numbers from the blog (15 sec), your take on what this means for solo builders (15 sec). Keep it conversational.
- 4
Edit with captions and side-by-side visuals20 min
Use CapCut or your editor of choice. Layer the screenshots under your voiceover. Add auto-captions. Use a split-screen moment when comparing the small model to the big ones.
- 5
Write the post copy and publish10 min
Firefox gets iOS ad blocking — what to make about it
Mozilla announced an ad blocker for Firefox on iOS (276 HN points) and a separate post urging people to stick with Firefox hit 277 points. This is not just browser news — it is content fuel for the 'take back your phone' niche. Privacy-focused creators can now demo Firefox iOS as a genuine Safari alternative with native ad blocking.
- Film a side-by-side: Safari with ads vs Firefox iOS without. The visual contrast is instant content.
- The 'Hang on to Your Firefox' post (277 points) signals audience appetite for browser-loyalty content — a surprisingly emotional topic.
- A Safari extension called Weedout that hides AI-labeled YouTube videos (36 HN points) is another angle: 'tools that give you back control of your feed.'
New tools that change what you can make today
Three GitHub repos worth knowing about if you make content with AI or teach others to.
- OpenMAIC (3,128 stars) is an open-source multi-agent classroom — think AI tutors debating each other. Edu-creators can demo this as 'I put 5 AI agents in a classroom and watched them teach.'
- video-use by browser-use (472 stars) lets coding agents edit video. Early but worth a 'first look' reel if you cover AI tools.
- minimind (1,005 stars) lets you train a 64M-parameter LLM from scratch in about 2 hours — pairs perfectly with the 'tiny model' content angle above.
Anthropic ships Fable and Mythos — what creators can do with it
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 (868 HN points, 816 comments — the most-discussed item in today's pack). For creators, the move matters if these models improve long-form writing, storytelling, or code generation you use in your workflow. The 816-comment thread is itself a content source: mine it for the best takes and turn them into a carousel or thread.
- The discussion thread has 816 comments — scan the top-voted ones for spicy takes you can quote in a 'what people are saying about the new Claude' post.
- If you use Claude for scripts or content drafts, do a before/after comparison with the new models and post the results.
- Pair with the AI-skeptic scorecard format: 'Anthropic said X, here is what actually shipped.'
Scripts to post today
TikTok90 Minutes to Beat a Billion-Dollar Model
“A solo developer just trained an AI in 90 minutes that outperforms models from billion-dollar labs.”
Here is the wildest thing I read today. One person, one laptop, 90 minutes of training. The result? A tiny transformer model that beats many large language models on the ARC benchmark — the test designed to measure actual reasoning. No massive data center. No billion-dollar budget. Just clever architecture choices and focused training. The blog post hit 556 points on Hacker News and the comments are full of researchers going 'wait, what?' This matters because it means the gap between solo builders and giant labs might be smaller than anyone thought. Link in bio if you want to try it yourself.
Why it works: Opens with a specific, surprising contrast (90 minutes vs billion dollars) that stops the scroll. The 556-point number adds credibility. Ends with a clear CTA.
LinkedInThe Platform Rug-Pull Post
“Google Play just blocked an open-source app's donation link. 818 developers are furious.”
Profiles trending right now
YouTubeSteal this: The 'small beats big with receipts' framing. Show exact benchmark numbers, not vague claims.
XSteal this: The 'digital archaeology' format — open an app's package contents and narrate what you find. Low effort, high curiosity.
Steal this: The prediction audit format. Pick any public figure's past claims, grade them, and publish the scorecard.
What to do today
- Post a 'draft or pass' reel using your niche (tools, strategies, platforms) while fantasy football search volume is at 5,000 on Google Trends — ride the format, swap the topic.
- Make a 5-slide Instagram carousel about the ChatGPT/LibreOffice discovery: hook, 3 evidence slides, question slide. Use screenshots from simonwillison.net.
- Write a LinkedIn post about a time a platform changed the rules on you. Reference the AnkiDroid (818 HN points) and AuroraStore (455 points) stories as your opening evidence.
- Record a 60-second TikTok or Reel explaining the small-transformer-beats-LLMs story. Use the hook: 'One person, 90 minutes, and a model that beats billion-dollar AI.'
- Publish an X thread grading 5 AI predictions from last year as pass or fail, inspired by the Dan Luu scorecard post (350 HN points, 414 comments). End by asking followers for their own grades.
Sources (14)
- I trained a small transformer in 1.5hrs and it beats many LLMs — Hacker News
- How accurate have Ed Zitron's AI skeptic predictions been? — Hacker News
- AnkiDroid: Google Play no longer allowing Open Collective donation link — Hacker News
- Play Store blocks AuroraStore, hurting GrapheneOS users — Hacker News
- The ChatGPT/Codex app bundles a full copy of LibreOffice — Hacker News
- Claude Fable 5.1 and Claude Mythos 5.1 — Hacker News
- Introducing Ad Blocker for Firefox on iOS — Hacker News
- Hang on to Your Firefox — Hacker News
- Show HN: Running 104GB Qwen3.8-Flash-Next on 48GB Mac with at ~12 tok/s — Hacker News
- Yahoo fantasy football — Google Trends US
