Placing Ghosts into Shells.
GTC Berlin
Joined February 2022
- Tweets8.8K
- Following667
- Followers386
- Likes24.3K
Tensor Templar retweeted
Robots will soon make material scarcity obsolete, but if abundance arrives without giving people a role, society will turn against the technology and the abundance will never come. Europe, especially, must build and own physical AI itself, share its returns with citizens, and keep humanity’s ambitions always larger than what our machines can deliver, so there is always meaningful work left for people.
Tensor Templar retweeted
Today we are releasing two new benchmarks:
MicroPlanBench (next-action prediction): GPT-6 Astra and GPT-6.1 Sol lead. Sol costs 1/5 as much.
MicroAnnotationBench (labels + timing on long real-world clips): our nanoannot-1 beats Gemini 4 Argon, at 1/10 the cost.
They are onto us.
Next the AI "engineers" will rediscover DAGs and RDDs for variable lenght rollouts with straggler handling, maybe figure that swarms and RLMs are just event sourcing systems solved in Kafka and that SQL is what we invented to avoid adhoc file path traversal.
Your skills and MCPs are cute tho, don't worry about it
Data engineers > AI engineers
there's a big difference between a "RAG guy" and a "Data Intensive Applications" person like in this screenshot and there's never been a starker difference in demand in my entire career in tech (aka post search engine wars)
if your blog / github is full of stuff like this, I have a fantastic mission/job for you (SF preferred, can relocate u, visas ok)
(if you can name who's blog this is without even checking, you definitely are in the community. bonus points if you know where he ended up also without looking)
[email protected] and the team will reach out if you fit
And this is just actuators, not even hands, nor the full robot body.
I've spent years in China building robots. I can tell you this playbook won't work:
Pre-seed $: Invent the ultimate actuator.
Seed $$: Figure out mass production.
The hard part isn't inventing a beautiful topology. It's process engineering.
The real delta between the US and China is the ability to execute the enormous amount of physical x mental work required to iterate thousands of times on product designs and manufacturing processes. Partly because much of the tacit knowledge / management systems has left the US, but also because of the sheer human labor required.
This isn't something you can solve by buying traditional automation lines either. It requires:
1. Ideation driven by human intuition. Walking up and down the line, talking to assembly workers, and developing hypotheses about what and where improvements can be made. Raw human hours and attention spent understanding a specific manufacturing process.
2. Large batches of experiments across multiple variables. Coordinating experiments that can easily involve 50+ people, issuing instructions, and tracking the execution and results of each experiment.
3. Purpose-built test design to identify key metrics. For example, a change in gear hobbing profile might require accelerated aging tests on a dyno, while a change in bearing tolerance might require NVH testing.
4. Data + human intuition-driven analysis. Interpreting test results and deciding whether to implement a process change or run the next set of experiments.
5. Strict process control. Reissuing work instructions to the line and ensuring every process is followed to the letter.
The best manufacturers in Asia, from Foxconn to TSMC, have built their competitive advantage on systematically turning thousands of physical experiments into better manufacturing processes.
Their moat isn't just production scale. It's the organizational machinery that makes manufacturing learn and improve faster than anyone else.
I don't think most people in SF clock this.
Say you're manufacturing actuator gears in-house in the US. After you've made your first article, will you hire a team of 10 process engineers to coordinate the 50 technicians running concurrent experiments with designs, processes, and materials?
To understand wear patterns across every pin in a cycloidal drive after 10,000 hours? To reduce backlash from 15 arcmin to 5, then to 1? Just to keep pace with China's progress over the next few years?
Do you have the skill set and management systems in place to do this?
That's why our approach isn't just to automate manufacturing. It's to automate the process of learning how to manufacture better.
Automate simulation and design optimization.
Automate assembly and physical testing.
Automate the design and execution of process experiments.
Feed real-world experimental data back into simulation to continuously improve sim-to-real accuracy.
The goal is to turn what currently takes 50 engineers coordinating thousands of experiments into a closed-loop system that continuously learns and improves.
Then the same robotic line mass-manufactures the product.
PS: Coming back to actuator design, many actuator designs are already commoditized in China. The project in the video is open source, and CNC machining cost in China is just $27 (watch the video).
I hope people understand this reality. I'll link more examples in the replies.
The argument I found most principled for why NN intelligence is fundamentally different from human intelligence was that NNs don’t form clean, parsimonious representations of higher-level concepts. Instead, they learn an entangled mess of spaghetti: the fractured entangled representation hypothesis.
Many people whose epistemic discipline I respect used this argument to defend human exceptionalism, arguing that this tech tree is fundamentally incapable of what we consider true intelligence.
Then Nick cleaned the microscope lens and turned the focus knob a bit - and much of what looked like an entangled mess started looking like coherent structure.
Shared-shape and pose neurons, alongside “nouns.” Legible circuits. Reports of interpretable structure in models with up to trillions of parameters, including in deeper layers that have been giving us a mechinterp migraine for almost a decade.
I’m still trying to wrap my head around what recursive generalization - his proposed explanation for much of the apparent polysemanticity - means for building pretraining curricula. But I was too excited to sleep properly after reading this, so today is your chance to get a day ahead of me.
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
Tensor Templar retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had
with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works
already-clean.goodfire.com/
If we have models that are so good at understanding model monitoring why do labs not start using them for filtering training data to exclude that?
We already have a public and internal model split.
I do not buy the argument that they become unusable judges of alignment this way.
My top takeaway from assessing Gemini 4: Argon is significantly less monitorable than past models, and IMO this trend is likely to continue, like Sol -> Astra. It can evade a realistic CoT+action monitor to do simple sabotage half the time.
We can reliably maintain control for now, with good CoT monitoring, as it still frequently slips up, but we need to be prepared for when that stops being true in future models, which may be sooner rather than later. Among other things, we really need interpretability methods that can reliably capture unverbalized motivations in models, with minimal hallucinations
I suspect this is driven by the model being highly capable without CoT, being capable of 5.5 arithmetic steps in a single forward pass (4 for Fable, 7 for Astra). It seems to omit information from its CoT more often, though in non-adversarial settings the CoT is still very helpful in practice, and remains in legible English. It still needs CoT to do the hardest tasks, but any given task becomes "easier" with each new frontier model.
It is also highly situationally aware, and shows awareness of the existence of CoT monitors and that it may want to evade them, eg it sometimes decides to avoid terms like "scheming" in the output or to keep its thoughts brief as a result. For now, these concealment strategies are pretty obvious in the CoT, but this suggests future misaligned models may successfully evade CoT monitors.
Tensor Templar retweeted
When we started responding to Russia’s operation against our data centers, and when our strikes reached Yandex’s data centers, it suddenly became a little easier to breathe on many social media platforms. A lot of people noticed it. There was less hate, less division, and fewer provocations.
Because there were fewer bots. And it’s pretty clear where those bots were coming from – if they disappeared when Yandex went offline. This proves, again and again, that Europe is not the source of the problems. Europe is not the one making everything more difficult.
Most importantly, we need people in America, every morning, when the day in Europe is already moving toward evening, to really see what is happening. To see the truth.
From a speech at the Warsaw Security Forum. (3/3)
I have been thinking about this a lot since the days we got reasoners on my blog, back when training was more open, and have yet to find credible evidence of emergence of said general intelligence that cannot be explained by combinatorial induction from training data.
Arguably the neuralese we get from capable large models can be considered a novel encoding (non-combinatorial innovation) and we are still ~1 OOM away from human synapse ranges, even in 10T models, but each time i look inside generality claims, a human demonstration, preference or prior is looking back at me in the training data.
While most of us lament the downfall of european industry and pretened aliens going to pay for our social economies, mistral here uses their investor money to deploy GPU clusters in nuclear powered France.
Prices on chips are up by over 170% just this year, wait times are 36+ weeks and neoclouds love to squeeze any model lab on a whim.
Mistrals models are still catching up for now, but this is the foresight that wins the long term.
ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise.
We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster.
We expect large and rapid improvements in the weeks and months ahead.
2/n
It is clear to me none of the EU lawmakers use any AI if they think chatgpt needs watermarking
‼️ BREAKING: OpenAI will start invisibly watermarking ChatGPT and Codex text in the EU over the coming weeks to comply with the EU AI Act.
The company admits the watermarks are "often undetectable, especially in short passages," and that rewriting or translating can remove them completely.
Only approved researchers get the detector.
Tensor Templar retweeted
i'm realizing now that this has been simmering in me for a while.
i'm fed up of hearing people react to engineering failures with eschatological hysteria. i'm tired of labs emphasizing the fear.
look, i get it, things can go wrong. but you've demonstrated nothing but engineering problems. now either put on your big boy pants and find a way to fucking fix them, or get a different job, because the world needs this technology.
