😸 Samsung's impressive tiny AI win

Welcome, humans.

Olympics memes are going viral on Sora, and they’re actually quite creative; for example, here’s Jesus competing in Olympics swimming…

There’s also animal versus human Olympics, the “un-Olympics” (or when doing the Oylmpics goes wrong), Olympics dishwashing, and even the smoking Olympics.

Now, for something serious, check out this piece from Matt Stoller explaining how China just threatened to pop the entire AI datacenter bubble by leveraging its monopoly on rare earth materials.

It announced that foreign suppliers now need Beijing's case-by-case approval to export products containing these materials (critical for semiconductors and AI chips), effectively giving China a kill switch on whether America's data center buildout can continue. The market lost ~$2T on Friday in response to US President Trump’s reaction to this, but futures trading rebounded yesterday when Trump said “it will all be fine.”

Naturally, we had to create “The Olympics of Tariffing” in response…

Here’s what happened in AI today:

1. We explain Samsung's new “tiny recursion model” and why it matters.
2. US projected to lead semiconductor investments by 2027.
3. Argentina partnered with OpenAI for $25B AI data center.
4. EU launched €1B AI adoption plan targeting 75% industry adoption by 2030.

Advertise in The Neuron here

P.S: We just published a fantastic interview with Mark Williams Cook, an SEO titan with LOTS of opinions and advice on how to operate an online business in the age of AI. Definitely check this one out if you missed it on Friday! Click below to watch.

Samsung's 7M-Parameter Model Just Beat Frontier LLMs. The Secret? Less is More.

For years, the AI race has been pretty straightforward: bigger models = better results. OpenAI throws billions of parameters at GPT-4. Google counters with Gemini. Everyone's scaling up like it's an arms race at the gym.

Then Samsung walks in with a model that's 10,000x smaller than these giants and proceeds to embarrass them on one of AI's hardest tests.

Meet the Tiny Recursive Model (TRM). At just 7 million parameters (the encoded numbers and calculations that make up how an AI model responds when prompted), it's beating DeepSeek R1 (671B params), Gemini 2.5 Pro, and o3-mini on ARC-AGI. ARC-AGI is the benchmark designed to test actual reasoning, not memorization. Think geometric puzzles that should be easy for humans but make AI models look silly.

Here's what TRM achieved:

• 44.6% accuracy on ARC-AGI-1 (vs. Gemini 2.5 Pro's 37% and DeepSeek R1's 15.8%).
• 7.8% on ARC-AGI-2 (the newer, harder version where most LLMs score under 5%).
• 87% on Sudoku-Extreme (up from 55% with previous methods).
• All with a model you could basically run on a laptop.

The trick? Instead of making the model massive, TRM uses recursive reasoning. It starts with an answer, then loops back through itself multiple times to refine that answer. Like drafting an essay, then editing it six times before submitting.

Samsung's researchers simplified an earlier approach called HRM (which we covered back in July!) by ditching its complex “hierarchical reasoning” theories that were proven to not be that relevant to its success. The TL;DR on that goes like this:

• The brain-inspired architecture wasn't the secret. A regular transformer (AI mechanism) dropped performance only 5%.
• The real driver was training on augmented versions of the exact test tasks.
• It used iterative refinement loops to improve answers through multiple attempts, memorizing task-specific patterns.

Samsung’s insight = for similar or better results, you just need two things: your current answer and a workspace for reasoning. Then recurse until you've polished it enough.

The catch: TRM trades parameters for passes. While the model itself is tiny, it needs multiple improvement steps (16 max) and took 3 days on 4×H100 GPUs to train (GPU = Graphics Processing Units, NVIDIA’s specialized computer chips originally built for gaming graphics but now the workhorses of AI training). So you're not saving compute per say… you're just spending it differently.

Why this matters: This isn't replacing ChatGPT anytime soon (it's specialized for grid puzzles, not chatting). But it proves that architecture and strategy can beat brute-force scaling. For you? If smaller models can match bigger ones on specific tasks, the AI tools you use at work could become faster, cheaper, and accessible without needing massive cloud infrastructure. As one researcher put it: “Recursion is the new scaling.”

In an industry obsessed with making everything bigger, Samsung just showed that sometimes, the smartest move is to go smaller and think deeper.

Check it out: Paper here, code here, video here (great breakdown by Matt Berman!).

FROM OUR PARTNERS

MASTER AI IN 16 HOURS AND BECOME IRREPLACEABLE BEFORE 2025 ENDS!

The spooky season is here—candies, costumes, and fear is everywhere.
But the real nightmare? It's not ghosts—it's job loss
Over 71% of people believe AI will take their jobs by 2025. The anxiety is real: Are you good enough? Fast enough? Smart enough?

But here's your treat this Halloween—a real solution to end the fear

Join the online 2-Day LIVE AI MASTERMIND by Outskill - a hands-on intensive training designed to make you an AI powered professional who can learn, earn and Build with AI.

Usually $395, but as a part of their halloween sale 🎃, you can get in for completely FREE!

Join in now, (we have limited free seats! )

Prompt Tip of the Day

Want to write your product descriptions like we write our Treats to Try tool suggestions? This simple prompt formula turns messy descriptions into clean, scannable one-liners that explain to your audience what your product actually does.

Here’s the structure: “Summarize this in one sentence. Format: Name + verb + what it does + funding (if mentioned). Use clear, non-technical language. Write in second person directly to the reader as 'you'.”

Then paste examples of the style you want, like:

• “Spiral automates 80% of your daily writing tasks while keeping your voice, tone, and style.”
• “Daydream is an e-commerce search engine that delivers personalized shopping results (raised $50M).”

The magic happens when you add these refinements:

• “Don't say AI-powered unless absolutely necessary” (consumers want outcomes, they don’t give a flying fig about “AI” or not).
• “Make it concrete—readers should instantly understand with a real-life example.”
• “Show don't tell.”

Those last two are key to making the description come to life. By adding “make it more concrete. Show, don’t tell” we’re able to improve 90% of bad, hard to understand sentences. Tell people what your product will do for them, not what buzzwords it hits for you.

Treats to Try

*Asterisk = from our partners (only the first one!). Advertise in The Neuron here.

1. CampusAI trains you on AI tools through courses and a 3D world, then gives you access to GPT, Midjourney, and 30+ models in one place—no juggling subscriptions. (raised $100M).
2. Prezent creates business presentations from your prompts and files in seconds, converts any deck to your company's brand template with a click, and has expert teams build polished presentations overnight. (raised $30M)
3. Foundation Health provides white-labeled infrastructure so you can launch your own digital pharmacy, telehealth service, and lab testing, handling prescription fulfillment, video doctor visits with 1,000+ clinicians, and at-home diagnostics all under your brand (raised $20M).
4. Quant from Sourcetable gives you Sourcetable stock analysis tools across 600+ exchanges to make smarter trades (read more)—free trial, then $20/month (Pro plan with 500 AI messages/month), or $200/month for unlimited.
5. Prohost handles your Airbnb guest messages, schedules cleanings, creates maintenance tickets, and suggests upsells to boost your revenue—free trial, then $30/month.
6. Intryc evaluates all your customer support interactions in real-time, coaches agents automatically, and integrates with 20+ helpdesk tools (raised $4.2M).
7. Attrove turns your scattered Slack messages, emails, and calendar events into a single dashboard of prioritized updates, so you can instantly see trending topics and team progress without digging through apps — free tier available, Pro plan with launch discount.

Around the Horn

Hilarious

• The US is projected to surpass China, Taiwan, and South Korea in semiconductor investments by 2027.
• Hollywood’s “above the line unions” (directors, actors, writers) are heading into contract negotiations, with the DGA expected to begin talks in November 2025 and SAG-AFTRA and WGA contracts expiring mid-2026; experts predict that AI actress Tilly Norwood and AI confusion will be key negotiating issues.
• Instagram head Adam Mosseri responded to MrBeast's warnings about AI's impact on creators by arguing that genAI will democratize creative production.
• The EU launched a comprehensive €1 billion AI support plan, aiming to increase AI adoption in key European industries to 75% by 2030.
• Argentina partnered with OpenAI to build a $25B, 500MW AI data center as Latin America's first Stargate project.
• OpenAI was forced to disclose Slack messages in a copyright lawsuit, and could next be used

FROM OUR PARTNERS

Save hours — in every app you use.

Typing wastes time. Flow makes voice the faster, smarter option everywhere.

Wispr Flow delivers no-edit confidence:

• 4× quicker than typing. Dictate emails, docs, Slack updates, or code.
  
• AI auto-edits. Filler words gone, grammar fixed, formatting intact.
  
• Works everywhere. From Slack, Notion, Gmail, and ChatGPT to Cursor, Windsurf, and iPhone.
  
• Smarter for teams. Snippet library drops in links, templates, or FAQs on cue.

Flow users save an average of 5–7 hours every week — full workdays back every month.

Give your hands a break ➜ start flowing for free today.

Monday Meme

A Cat’s Commentary