Pinned Tweet
Climate change evidence: How do we know? climate.nasa.gov/evidence/#.… thanks @NASA
So nvidia made $96B in last quarter.
Didn’t cure the cancer, but made $663 million in profit every single day.
🤯
35
Following up on this - is AI inference actually a good business?
My take was: inference brings in the customers, post-training is where the money is.
So I went through 10 recent podcast interviews with AI infra execs - Modal, Fireworks, Nebius, Crusoe, CoreWeave, OpenRouter, Cerebras. Turns out I was only half right.
Nobody calls raw inference very profitable. Selling tokens for dollars has no switching cost - it's a race to the bottom. The margins people mention are around 25-40%, and a good part of that is just the GPU shortage. Everyone is short on supply right now.
Customization is what keeps customers - fine-tuning, post-training, RL on your own data. Fireworks says most of their tokens already come from customized models.
But the money is still billed as serving. You customize once, then you pay for inference forever.
So inference is the cash register, customization is the lock-in.
And the fattest margins in the whole stack? Memory makers, at ~80%.
nitter.cf/YingjunWu/status/21212…
I've recently talked to quite a few people recently who believe they can just rent B300 nodes, install @vllm_project / @sgl_project, plug into @OpenRouter, and start printing money.
I think they seriously underestimate how hard the inference business has become.
A few facts everyone should know:
- Getting listed is extremely hard now! OpenRouter reportedly gets 40 provider applications every day and only approves 1. It could take months just to get listed!
- Competition is brutal! There are already 80+ providers serving @deepseek_ai models. Unless you have done extreme optimizations, you won't get any meaningful traffic!
- Some competitors aren't even trying to make money! Some providers intentionally price below cost to attract traffic. They treat the losses as marketing spend, and no one can even compete the price!
I think the era of simply hosting open-source models and selling tokens is coming to an end, especially for new entrants.
The real opportunity is moving up the stack. Sell solutions, not tokens.
1
2
96
The idea, that you can pinpoint GPU model to the model output is 🤯
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
19
Looks like November 19 is going to have two kinds of people.
Those playing GTA 6.
And those watching the Anthropic IPO.
Btw, it's also International Men's Day.
1
2
178
Can we reboot Earth? // There is a glitch in the matrix
Gen Z fueling tobacco sales as wild claims cigarettes are 'healthier' than vaping go viral: 'plant-based alternative' trib.al/tsOeuG2
Readers added context they thought people might want to know
Cigarette volumes in the US continue declining (offset only by price hikes) and youth use is at historic lows per federal data; company growth is driven by smoke-free products, not traditional cigarettes or Gen Z anecdotes.
fda.gov/tobacco-produc…
tobaccoinsider.com/usa-cigarette-…
pmi.com/our-progress
29
A couple of years ago I was convinced the market for small language models would be huge. I was wrong.
But about a year ago I noticed something else - most ASR (speech-to-text) models really struggle with names and technical terms. The ones with an LLM decoder can sometimes guess a fix. Almost all the others won't know how to write "Kubernetes" if it was never in their training data.
I hit this while building a "library" tool for the podcasts I listen to. It pulled me into a deep dive on how LLMs work, and ASR models especially. After a bunch of experiments, including fine-tuning @NVIDIAAI Nemotron, I decided to build my own model. Absolutely from scratch.
It was a hell of a ride and I learned a lot. Now I can say from personal experience - it's really hard. My model wasn't amazing. But it was okay.
Mevine 0.1 is a 107M-parameter Conformer RNN-T. No pretrained encoder, no distillation, random init. 5,167 hours of English audio, 15.7 hours on a single H100. Average WER is 13.5% on the Open ASR Leaderboard sets.
huggingface.co/theaviv/mevin…
Then I built freeaudiotranscription.co, so anyone with a browser can enjoy Mevine ;)
I went for the browser on purpose. I quantized the encoder to INT4 and exported it to ONNX - 429 MB down to 94 MB, with WER basically unchanged. It runs on WebGPU, right on your own machine. The idea was to make it work on any computer, no install, no GPU server.
huggingface.co/theaviv/mevin…
I'm into inference and optimization, and I've spent years on data pipelines, so of course I added anonymous telemetry. No IPs, no personal details, just things like RTFx and model load time.
Shipped it, posted a couple of links and moved on to a new work project.
A couple of days ago I wondered - does anyone actually use it?
Apparently, yes. Since mid-August:
988 files transcribed, about 225 hours of audio.
People came from 34 countries, and 26 of them actually transcribed something.
121 files were longer than 30 minutes. The longest one was 2h23m.
Median speed is 34x real time, in a browser tab. All 225 hours took less than 9 hours of compute.
Most files came from the Philippines. Most hours came from the UK - a few people there are clearly running whole recordings through it.
Not all pretty though. About 1 in 8 events is a crash, mostly browsers without WebGPU and audio that won't decode. So that's next.
1
70
Hacktoberfest recap for Klag. A bit overdue.
I maintain Klag, a Kafka consumer lag exporter
github.com/themoah/klag
On 21 Sep I put up a contributor board:
github.com/themoah/klag/issu…
Closed it as completed on 3 Nov.
7 people new to the project got their work merged. 13 PRs merged, out of 17 opened. Nobody who already contributed to Klag came back for this one. Every merged contributor was new.
There is a StatsD/DogStatsD reporter now. Also partial-cycle metric cleanup, Helm disruption budgets, a bunch of MCP hardening (JVM -D settings, rejecting bad JSON types, Origin/Content-Type checks, blocking multipart uploads, locale-safe labels), CI on Java 25 LTS. And docs for Confluent Cloud, a kafka-lag-exporter cutover, Grafana/PromQL polish and an MCP/Prometheus checklist.
One person landed five of those merged PRs. Thanks to everyone who contributed.
A few issues are still open if you want to jump in - MSK IAM, topic filters, a Datadog dashboard and a local integration stack.
github.com/themoah/klag
2
1
173
Yey, finally got the blue mark ✔️
Took a bit more than a week of "under review".
First perk I'm going to abuse - longer posts. Kafka, Klag and tiny ML models, without squeezing everything into 280 chars.
First long one goes out later today ;)
1
35
Someone tried to sneak a remote-code loader (Crypto stealer) into my open-source project.
I maintain Klag, a Apache Kafka consumer lag exporter
github.com/themoah/klag
They didn't touch the Java exporter. They came dressed as documentation.
A first-time contributor opened two PRs on 13 Oct. One would merge CLAUDE.md into AGENTS.md. The other added a label cardinality guide I'd actually asked for. They looked like a gift. Hacktoberfest energy. "Docs-only." website/package.json was edited so `npm run dev`, `start`, and `build` run `node init && …`. If you maintain a docs site, this is the line a tired review skips. The new website/init file looks like it only creates a `.tmp` folder. That's the rubber-stamp part. Harmless. Helpful, even. Scroll to the end of the file. After a long run of tabs and spaces: a layered, obfuscated JavaScript dropper. The same class of hidden tail is tacked onto website/astro.config.mjs, so merely loading the docs-site config would run it too.
I unpacked it statically. I did not run it. It's a remote-code loader. It queries Ethereum for transactions from a hardcoded sender, then reads two IPv4 addresses out of a transaction's `to` field. Those IPs serve a second-stage script, which it XOR-decodes and runs in a detached hidden `node` process. Blockchain as a dead-drop: the C2 address can rotate without anyone editing the PR. You merge once; they retarget later. This was an attempt.
The PRs were closed without merge. The pages — and the account that opened them — are gone from GitHub now. Nobody running Klag was infected.
Attacker has already deleted the PRs, sharing screenshots:
2
3
157
I've spent over 7 years at cybersecurity companies as Data & ML Engineer and it almost slipped through.
If you maintain a docs site: read the tail of "docs-only" files — whitespace is a hiding place — and diff npm script hooks like production code.
Share this so other maintainers see the trick.
If you work with Kafka, star or watch Klag: github.com/themoah/klag
1
19
What a day for cybersecurity.. first OpenAI files 948 patches, then vuln from 2016 OpenSSH on musl is being released.
31
First thought while reading news on hacked fridges in Belgium and Sweden:
The son of Anton is back.
40
Reading about OpenCode security incident.
"An intruder reached a legacy system at OpenCode".
Afaik OpenCode was released in June 2025.
That ain't legacy. That's the price of rushing
thehackernews.com/2026/10/op…
1
48
Should we expect grok 4.8 soon @SpaceXSI ? Looks like 4.7 was dumbed down / quantized
83
Am I too late to the party and Jev-alike models are the yesterday new, or anyone is still interested in them?
I'm happy to announce that I'm releasing Lev1, based on @IFM_AI K2 Horizon - 3.7B model.
themoah.medium.com/introduci…
1
77
It's fast, ~28ms median response time.
Great - 68.8 on Tools & Automation benchmark (Decision index v0.2.1), on par with 27B (8x bigger) models.
Apache-2.0 licensed.
And I'd really appreciate the feedback. Please don't hesitate to message me or open a discussion on @huggingface
1
17
Additionally I'm releasing the dataset that I've build for post-training - huggingface.co/datasets/thea…
13













