Sam Lo

Open Source - Robotic - Local LLMs

[Sam Lo @SamLoAix]

Can open-source models kill OpenAI and Anthropic? is the wrong question. The revenue data says something more interesting: open weights aren't killing them. They're reshaping them.

OpenAI: $20B ARR at end-2025 → $40B+ run rate this week. Anthropic: $9B → $30B+ in five months, with its first profitable quarter in Q2 — training costs included.

The pressure is real. US companies' token share on Chinese models via OpenRouter: above 30% every week since February, peaking at 46%. A year ago: 11%.

The response: OpenAI cut Luna 80% and Terra 20%. Altman says he's "happy to deliver at a quarter of the price." Anthropic shipped Opus 5 at Opus 4.8's sticker. Google keeps undercutting both.

Here's the mechanism people miss: OpenAI's revenue tracks compute, not price. Revenue grew 10x on 9.5x compute from 2023-25. Price cuts at the commodity tier buy volume; the frontier tier still holds its premium.

What would actually break them? The frontier gap closing. GLM-5.3 just went 4.6 → 28.3 on Terminal-Bench 3.0. Kimi K3 scores above Opus 4.8 on GDPval. Brookings puts Chinese models 6-9 months behind(I think it's less than 6 months).

The counter-case exists: Cathie Wood argues open weights make the frontier labs richer — cheap models grow the market, and the hard 5% still needs the best.

The real threat isn't open weights. It's routing. "When a task doesn't need the best model, teams route to the cheapest good-enough one" — that's Vercel's data, and ~80% of workloads don't need frontier.

👀

GLM-5.3 weights (two weeks), Anthropic's reported $2T October IPO, OpenAI's first device (H2 2026), and whether Sonnet 5's +50% September price increase survives contact with the market.

As a result, competition benefits the consumer.

[Sam Lo @SamLoAix]

Whether you like him or not is a separate matter, but I watched Bill Gates’ video that has been circulating in recent days. Unfortunately, I agree that a significant part of what he says about artificial intelligence is correct.

I agree with his assessment that humanity is largely unprepared for the changes AI will bring, that governments are not doing enough to prepare for what lies ahead, and that we are likely to face a major wave of unemployment.

Of course, it won’t happen overnight. But over the coming years, I believe we will gradually begin to feel its impact, step by step. Even in the United States, unemployment is projected by some to rise to between 5% and 10% within the next five years. A 10% unemployment rate would be an extraordinarily high figure for the United States. I can hardly imagine what the consequences could be for other countries.

We need to take this seriously and prepare for it as individuals, while governments must also play an active role in that preparation. Serious regulatory measures are essential. A fundamental transformation of the system seems inevitable.

[Sam Lo @SamLoAix]

Bosch has signed a contract with the British robotics company Humanoid to manufacture the HMND 01 humanoid robot for the European market. This is seen as the first serious move into humanoid robotics by a major European manufacturer.

[Sam Lo @SamLoAix]

Does it surpass Opus 5 in some metrics? Have they achieved the impossible? Will the next announcement be an open-source model that also surpasses Fable?

[Tencent Hy @TencentHunyuan]

🚀 Hy4 preview is here.

770B, 49B active, 1M context.
Built for productivity.
Open source frontier.
Consistent affordable price.
Use it. Tell us what breaks.

More on Hy blog：https://t.co/rbl1IWRk3C
HuggingFace：https://t.co/mE9wevH5XR
Github：https://t.co/pyl9zckpoL

[Sam Lo @SamLoAix]

While the model size has been reduced by approximately 60% compared to the GLM 5.2, its performance has increased. Fantastic. This shows us something.

Deepseek, Qwen, and GLM series offer much smaller models that rival the 1T+ models, but with the same or higher performance, all as open source.

The question on everyone's mind is, will models become even smaller while still providing higher performance? And how would that affect the argument that we need very high VRAM?

[Z.ai @Zai_org]

Introducing GLM-5.3-Flash

- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips

Blog: https://t.co/tzOmB7gdZP

Available now across all official platforms:

Weights: https://t.co/9LRMahY9Wa
API: https://t.co/VcaQnzYmS9
Coding Plan: https://t.co/Nk8Y98HNhU
ZCode: https://t.co/Peepqv4XSx
Chat: https://t.co/WCqWT0qCQb
AutoClaw: https://t.co/aGEG5HqTTb

[Sam Lo @SamLoAix]

For AI, pay close attention to bandwidth and choose the highest possible one. Nowadays, this is 1tb/s+. Otherwise, a 5-minute task might take 45 minutes.

[Sam Lo @SamLoAix]

Apple shipped its first 2nm chip inside a $899 box today. The Mac mini with M6 isn't a spec bump - it's Apple betting that the most important Al hardware of the next few years sits on your desk, not in a data center.

M6: 12-core CPU, 12-core GPU with neural accelerators in every core, dual 16-core Neural Engine, 170GB/s unified memory. Apple claims 4x the Al performance of the M4 Mac mini.

But the real story is what this unlocks.
LLM speed is memory-bandwidth-bound, not FLOPs-bound. 64GB of unified memory at 307GB/s means running serious local models on a $1,699 box that draws a fraction of a workstation's power.

This is the "always-on agentic device" play. Tim Cook said it in July: customers are using Mac minis as agentic Al platforms. 24/7, isolated, under your control. Demand is so strong some configs have 3-month waits.

Apple builds its foundation models on Google's Gemini, routes heavy reasoning to Nvidia GPUs in Google Cloud - and keeps the one layer that matters: the trusted surface where an agent sees context and takes action. The model is a supplier.

There's a catch hiding in the price. M6 starts at $899, up from $599 two months ago. Al data centers ate the DRAM supply. Memory is now the bottleneck on local Al - not compute. Apple even moved a 20B-parameter model into flash to route around it.

2nm + per-watt efficiency + Thunderbolt 5 clustering means "your own Al cluster" is now a realistic desk setup. The economics of inference are splitting: $0 marginal cost locally, metered cost in the cloud.

The next frontier isn't a bigger model. It's who owns the surface where the agent acts. Apple just made its strongest move yet to own that.

[Sam Lo @SamLoAix]

@MiaAI_lab 4 DGX Sparks consume approximately 3x more electricity than 1 M5 Studio 512GB. This means you'll pay three times more for electricity.

[Sam Lo @SamLoAix]

A 4-bit model just beat its own full-precision parent on 7 of 9 benchmarks.

That shouldn't happen. Smaller plus quantized usually means worse. Unless the recovery step was the problem all along.

The work: arXiv 2608.20953. GPT-OSS 120B, compressed to 60B, quantized to MXFP4.

The standard fix for quantization damage, QAT, is fragile. Converges slowly. Push training past its peak and the model collapses — so shipping safely means hand-tuned early stopping. That's a standing liability in production.

Their move: skip the recovered bfloat16 checkpoint as teacher. Distill the 4-bit student straight from the original, uncompressed 120B model. Teacher and student don't even share an architecture — the teacher's output distribution transfers anyway.

Result: the 4-bit 60B model beats its bfloat16 source on 7 of 9 benchmarks. +7.4 on long-context reasoning, +5.6 on competition math. It matches the full 120B teacher on LiveCodeBench, at 4x less weight memory and half the parameters.

The recipe also peaks ~7x faster than QAT and holds stable — no early stopping roulette.

These are the team's own measurements. No independent eval yet, and the head-to-head against QAD, the other distillation baseline, wasn't run. Worth treating as promising, not settled.

[Sam Lo @SamLoAix]

Apple shipped its first 2nm chip inside a $899 box today. The Mac mini with M6 isn't a spec bump - it's Apple betting that the most important Al hardware of the next few years sits on your desk, not in a data center.

M6: 12-core CPU, 12-core GPU with neural accelerators in every core, dual 16-core Neural Engine, 170GB/s unified memory. Apple claims 4x the Al performance of the M4 Mac mini.

But the real story is what this unlocks.
LLM speed is memory-bandwidth-bound, not FLOPs-bound. 64GB of unified memory at 307GB/s means running serious local models on a $1,699 box that draws a fraction of a workstation's power.

This is the "always-on agentic device" play. Tim Cook said it in July: customers are using Mac minis as agentic Al platforms. 24/7, isolated, under your control. Demand is so strong some configs have 3-month waits.

Apple builds its foundation models on Google's Gemini, routes heavy reasoning to Nvidia GPUs in Google Cloud - and keeps the one layer that matters: the trusted surface where an agent sees context and takes action. The model is a supplier.

There's a catch hiding in the price. M6 starts at $899, up from $599 two months ago. Al data centers ate the DRAM supply. Memory is now the bottleneck on local Al - not compute. Apple even moved a 20B-parameter model into flash to route around it.

2nm + per-watt efficiency + Thunderbolt 5 clustering means "your own Al cluster" is now a realistic desk setup. The economics of inference are splitting: $0 marginal cost locally, metered cost in the cloud.

The next frontier isn't a bigger model. It's who owns the surface where the agent acts. Apple just made its strongest move yet to own that.

[Sam Lo @SamLoAix]

Still arguing about which model is smarter?

Wrong question.

The harness — the tools, the context handling, the verification loop wrapped around the model — moves scores 2-5x more than the model itself does.

Same model. Four different harnesses. Four different answers.

Claude Opus 4.5 ran SWE-bench Pro through three agent systems: 50.2%, 52%, 55.4%. The weights never changed. Only the scaffolding around them did.

That spread — five points with the model pinned — is the same size as the gap between adjacent rows on today's leaderboards. Which means the rank order you're reading might just be whoever wrote a better scaffold.

The controlled numbers are worse.

Swap the harness: pass@1 moves 8.5 to 13 points. Swap the model: 2.5 to 5 points. A single tool addition flipped the ordering between two frontier models. Harness effects routinely dwarf what papers call "meaningful model advances."

The cost side is where it gets absurd.

Databricks ran the same Opus 4.8 through three harnesses: 742K tokens per task. 1.2M. 237K. Same success rate. Same model. A 3-5x bill difference decided by the wrapper, not the weights.

NVIDIA's NOOA hit 82.2% on SWE-bench Verified — above the published SOTA — with half the tokens and 56% fewer LLM calls. Not a better model. A better harness.

And nobody verifies any of this.

Of 100 models on llm-stats, exactly one carries an independent verification badge. Vendors report their own scores, measured under their own chosen scaffold. Anthropic's Opus 4.8: 69.2% on its own harness. 51.9% under a standardized one. Same model family, 17 points apart.

So what do you actually do with a leaderboard?

Use it as a tier filter, not a ranking. It tells you who's in the frontier club. It doesn't tell you who wins inside it.

Then measure the harness yourself — on your own tasks, with your own tools. That number is the only one that decides anything.

The model is the engine. The harness is the car. We spent years comparing engines while the car was deciding the race.

[Sam Lo @SamLoAix]

Xiaomi, Genkai 0100 (XRING 0100 / 0100).

Xiaomi has unveiled its own in-house, high-memory-bandwidth AI accelerator chip.

They claim it can deliver up to 1.22 TB/s of memory bandwidth. If true, that would be faster(or around the same) than the bandwidth expected from the Mac Studio scheduled to launch later this year.

[Sam Lo @SamLoAix]

The countries that spent the most on ai in 2025 were the United States, China, the United Kingdom, France, and Canada. I expect many other countries to join these countries as well.

[Sam Lo @SamLoAix]

When it comes to artificial intelligence and the future, the first thing robots will need in order to work is electricity.

The cost of that electricity matters a great deal. Look at China, the United States, and Europe, and the gap is striking.

Today, China can generate power for about 3 cents per kilowatt-hour, while Europe sits around 5 cents and the United States around 6.9 cents.

Whoever drives the cost of electricity generation down the most will hold one of the greatest advantages of the coming era.

[Sam Lo @SamLoAix]

AI will make our lives much easier and reduce costs. The question is, how many people will have high purchasing power in the future?

[Sam Lo @SamLoAix]

A frontier-class model just showed up on OpenRouter. Free. 1M context. Video input. Nobody knows who built it.

It's called Ox Alpha (stealth/ox-alpha). A reasoning model aimed at coding and long agent runs: 1,048,576-token context, 131K max output, text/image/video in. Released Aug 20. $0 in, $0 out.

The mystery has a shape. This is the fifth anonymous release in six months. The previous four all got claimed by Chinese labs - Xiaomi MiMo, Zhipu GLM, Ant Lingxi, Meituan LongCat. Community guesses for Ox Alpha: a GLM-5.3 variant, or Xiaomi MiMo-V3. All guesses.

Why ship a model wearing a mask? Anonymity strips brand bias from evals. A lab gets frontier-scale traffic from real production workloads with zero PR risk. Claude Code and Hermes Agent were already pushing tens of billions of tokens through it within days.

What early users report: strong on creative and open-ended reasoning - some say it beats Kimi K3 there. Weak at front-end/CSS. Visual reasoning mediocre. Censorship behavior inconsistent: some get detailed Tiananmen answers, others get blanks.

The catch: zero published benchmarks. "Frontier" is the operator's own word. And the privacy story wobbled - OpenCode tweeted "zero data retention," while OpenRouter's listing says prompts ARE retained, just not used for training.

The play is simple. Use the free week to audition it on your own evals. Keep a fallback model underneath - the free tier may not survive the week and the real price is unknown. Don't send anything you wouldn't want logged by an anonymous operator.

[Sam Lo @SamLoAix]

Deepseek v4 flash is now an Opus 4.8. They're achieving something unbelievable. The model is competing against models that weigh much more than itself. Congratulations.

Thanks also for the Vision support.

[Sam Lo @SamLoAix]

Moderna just ran the first Phase 3 of a personalized cancer vaccine and hit both endpoints.

The interesting part isn't the biology. It's that this was never only a drug problem - it was a manufacturing problem Al quietly solved.

Every patient's tumor has a unique mutation fingerprint. Moderna sequences tumor and blood, and an Al pipeline picks up to 34 neoantigens most likely to trigger an immune response.

How good is that prediction? A 2026 lab validation of neoepitope prediction tools found only 22% of predicted neoantigens actually get presented on HLA molecules.

22%. And it still worked.

Because 34 candidates is a redundancy strategy. The immune system needs enough true positives, not a perfect list. The Al doesn't have to be right about everything - just good enough, fast enough.

Speed is where Al earns its keep. In 2019, sample-to-dose took 50-60 days. Now it's about six weeks, per Moderna. They only moved into fast-progressing metastatic melanoma after proving they could reliably deliver in that window.

Then the unglamorous half: Maestro, an Al scheduler that fits thousands of individual patient batches onto manufacturing capacity in real time. One medicine, one patient, thousands of parallel batches - treated as an operations problem, solved like one.

The numbers so far: Phase 2b five-year data showed a 49% lower risk of recurrence or death (HR 0.51) vs Keytruda alone, per Merck. Yesterday's Phase 3 topline met both primary and key secondary endpoints.

Behind all of it sits Lucy - a closed-loop ML platform running hundreds of thousands of experiment iterations per cycle, up to 1,000 barcoded candidates tested in a single animal, built with OpenAl.

Bancel's own framing: the moat isn't the mRNA. It's learning velocity.

That's the honest version of Al in drug discovery. Not a magic molecule generator - a good-enough selector plus a world-class scheduler that turned "one drug per patient" from science fiction into a Phase 3 product.

[Sam Lo @SamLoAix]

Goldman Sachs research finds that AI is eliminating approximately 16,000 U.S. jobs each month.

If the number doesn't increase, I don't think it's a big problem.

[Sam Lo @SamLoAix]

New midsize Qwen 3.8 model coming next week (hopefully) according to community manager!

If this news is true, it will be a good balance between small and large models. I'm eagerly awaiting its performance.