I’ve worked that heatshield my entire adult life. Peak career moment today
To go from little 7 tile patches to a full reusable ship. Absolute marathon
I’m so grateful I got to do something useful, that I was given the chance
I don’t really have many good memories from this experience.
It was so difficult. It almost looks easy now, but none of it came easily
Please do something amazing with this rocket. This tech is such a gift, a love letter to the species
me, MBA2 at UVA Darden, will literally drop out for a role at a portco. MBAs are cooked. Already put in speedrun talent network app. I'm the only M7 mba out there with a GitHub like this. Even made interactive games for all HuggingFace courses for our class to learn robotics ASAP + bot a Microduck robot.
https://t.co/v75Kdo5QcT
@justmazer@speedrun
You might have 1–5 years to contribute intellectually to the world before AGI. What should you do with them?
Excellent advice from @jaschasd (Anthropic) at Stanford. Skip Netflix tonight and watch this instead.
.@sama was almost peer-pressured into being a banker. Students are funneled down career paths before they’ve explored their options. With everything going on in tech, that doesn't sit right. I've been in your shoes, so I'm opening up off-cycle recruiting. https://t.co/7W64I5Z9M1
i got curious about what dario (anthropic CEO) is predicting next, so i went through all his recent essays and interviews
if he keeps being correct, here's what our future timeline looks like:
> 2026: someone builds a $1B company with just 1 employee
> late 2026-early 2027: AI matches or beats nobel prize winners in most fields, from biology to math
> 2027: chinese AI catches up to Mythos, the model anthropic kept from the public because it's so good at hacking
> 2027-2028: AI gets better than every human at almost everything, with millions of copies running 10-100x faster than us
> 2027-2028: AI starts building the next generation of AI on its own
> 2028: AI is using 100 gigawatts of power, about a fifth of all the electricity america uses
> by 2030: AI disrupts half of all entry-level office jobs
> before 2030: AI brings in trillions of dollars a year, and the economy grows 10-20% a year
> early 2030s: cancer deaths drop 95%, alzheimer's becomes preventable, and nearly every infectious disease can be prevented or treated
> early 2030s: most mental illness, like depression and PTSD, becomes curable
> 2030s: sub-saharan africa gets as rich per person as china is today
> 2030s: the human lifespan heads toward 150
Bir fizikçi itirafı:Doktora tezimde 1 yılımı ve~500 sayfa elle yaptığım kuantum gravite hesabını Opus 5.5 makalesi ve GitHub reposu dahil 10 saatte yaptı.😱 Üstelik hataları da buldu.Neyse ki sonuç değişmedi. Akademide “emek” bundan sonra neyi ifade edecek?https://t.co/fx3CgyYFnQ
Had early access to Gemini 4 Argon - so I asked for slides on the AI power bottleneck.
It synthesized 8 years of 10-K filings, FERC queue data across all 6 US grid operators, 17 gigawatt cluster buildouts, and a decade of GPU thermal specs.
And I got this interactive deck 🤯
Coming back from the dead is arguably one of if not the very best masculine experiences for confidence
I have spoken to some very high level operators who have been destroyed and zero'd out in their mid thirties by bankruptcy, divorce getting ugly and a multitude of other things and those individuals who eat those shots and came back to the arena swinging and rebuilt themselves from basically nothing; everyone of those motherfuckers has that untouchable it factor swagger and its so potent
Those moments where you're down and out, took a risk which was 95% of NW; got clipped and now you barely have money for rent without hopping on uber for 30 hours a week for the next 4 weeks, or maybe you even go into heavy debt and get clipped
Those are the moments that forge the spirit and make the man in complete totality, if a man can connect to his vision for himself and revive from those angles from the dead he will know a strength and a power like no other
Fear is such a neurotic mistress, its so easy to be consumed by fear when everything is going perfectly and it can paralyse you even when things are great, but then once you get destroyed into the dirt and all your nightmares are at the door, all of them have transformed and become real now it is strangely at this very point when they have no power over you, the point where you're in the most danger you care the very least as you have gained strength from engaging and welcoming your shadows, once you welcome something and you are truly not bothered by it; you swallow its soul and it animates you into correct posture and life force and you can utilise that as an additional limb of some sort
It's never getting decimated that really deconstructs men, you will always live in a glass house afraid of the wind when you have never realised you can survive the coldest nights and its genuinely that realisation that is the foundation of becoming and growing beyond who you actually are
Unless you fail truly at some point, you will never develop the courage to be who you truly could be, as the protection of your little ego will prevent you from truly taking on the almighty behemoth challenges
Failing brutally atleast once destroys your ego and liberates you, it almost acts as an additional health bar
Always the dudes who have done the least who are the most stressed and the men who are the most calm who have been in the war the longest
Otherwise you will live to protect your illusions forever
theres a method you learn in YCombinator to compress a month of GTM into a few days.
The goal is to figure out if customers will actually pay, or if they're just being polite.
Here's how:
1. Ask Claude to analyze competitor websites, customer reviews, Reddit complaints, and pricing pages.
2. Identify problems people already spend money to solve, not just problems they complain about.
3. Talk to 10 potential customers as soon as possible.
Phone them, email them, contact them on LinkedIn.
Never ask "would you use this?"
Ask instead:
→ How do you solve this today?
→ How much does it cost you?
→ What happens if you do nothing?
→ When did you last pay to fix this?
→ Who controls the budget?
4. Watch for red flags:
"Cool idea" = no signal.
"Keep me posted" = no urgency.
"Can you send pricing?" = mild interest.
"Can we start a paid pilot ?" = real intent.
5. Put a price on the table.
"We can solve X for $Y/month. Want to get started this month?"
Then shut up.
If they say no, ask why. Don't discount immediately.
6. Feed every conversation back into Claude. Ask it to challenge your assumptions and identify what would actually make someone buy.
7. Use @gojiberryai to find more prospects showing relevant buying signals. Test your hypothesis against the market.
The point is NOT to get people to validate your idea.
It's to find evidence that they're willing to pay for it.
Alexandr Wang on what matters in an age of abundant intelligence:
“The scarce resource isn’t going to be intelligence or agency. I really think it’s going to be vision and ambition.”
“Do you have a clear view of what you want the world to look like in the future? What is the one way in which you wanna put your finger on the scale for how the future of the world will develop?”
“AI will make that easier... But the flip side of that is then all of a sudden you can dream bigger.”
geoffrey hinton mass-quit google in 2023 and mass-told everyone he was scared of what he built. nobody asked the right question.
not "is AI dangerous." the right question is "why does backpropagation work at all."
hinton couldn't answer it. neither could lecun. neither could sutskever when he left openai. the three founders of deep learning built a $4.6 trillion industry on an algorithm none of them could mathematically prove should work.
until last month.
a physicist at ETH zurich published a 41-page proof that backpropagation is a special case of the second law of thermodynamics. not an analogy. a derivation.
the math:
thermodynamics: dS/dt ≥ 0. entropy of a closed system never decreases.
backpropagation: dL/dt ≤ 0. loss of a trained network never increases.
one is entropy going up. the other is loss going down. they are the same equation with a sign flip.
the proof shows that gradient descent on a neural network is mathematically identical to a physical system falling toward thermal equilibrium. every weight update is a molecule finding its lowest energy state.
this is why transformers scale. this is why chinchilla scaling laws look like thermodynamic equations. this is why nobody could prove convergence guarantees - they were using the wrong field. the answer was never in computer science. it was in statistical mechanics.
now watch what this means for the frontier labs [Fig. 1] -
openai trained GPT-4 at an estimated mass-energy cost of 50,000 MWh. anthropic trained claude 3.5 at approximately 31,000 MWh. google trained gemini ultra at over 60,000 MWh.
the ETH proof predicts a theoretical minimum energy for each model's capability level. the equation:
E_min = kT × N × ln(1/ε)
k = boltzmann constant. T = operating temperature. N = parameters. ε = target loss.
GPT-4's actual energy usage: 17x above the theoretical minimum.
claude: 12x above minimum.
gemini: 23x above minimum.
anthropic is closest to the thermodynamic floor. not by accident. dario amodei has 3 physicists on the training infrastructure team. openai has zero.
the companies that understand this equation will compress 3 years of scaling into 6 months. the companies that don't will spend $10 billion on training runs that a physicist could tell them are 20x less efficient than possible.
hinton understood the gradient. he never understood the entropy.
the network doesn't learn. it cools.
save this
Superintelligent Economic Zone (SIEZ)
/noun/
A designated region where superintelligence acts as the economy's operating system, autonomously orchestrating capital, energy, compute, robotics, infrastructure, and human activity.
SI coordinates production, trade, logistics, and innovation in real time, creating a continuously evolving, self-optimizing economy.
See also: Agentic Economies, Special Economic Zones.
Great to be on this Moonshots AMA as a UVA Darden student. @DaveBlundin’s advice to MBAs: apply management skills to agent swarms ASAP.
We’ll all manage 100 to 1,000 agents within a year, with IQs of ~150, each working 10-100x more efficiently than a person. 1000 agents = 100k employees.
"Solve Everything," by @alexwg and @PeterDiamandis predicts math and software solved first (2026–27), then chemistry, materials and biology (2028–31). On cue, OpenAI published results on 372 open math problems this week from one prompt given to one agent.
Dave's challenge: if you were handed a million brilliant employees, could you build a company with them?
Compute and power are still constraints, but deciding where to point this firehose of intelligence is a management problem as well as a new golden age of entrepreneurship.
If you're an MBA heading into tech, start now. Thanks to the Moonshots crew, and thanks for the Darden shout-out, Dave!
Seeing similar at top MBA programs. Currently 2nd year at UVA Darden. I'm fully Claude Code since Jan, Wispr Flow 2M+ words. Nearly 100% use AI but 70% of them use only chat or Cowork. 23% use CC/codex by now. <2% are SERIOUSLY preparing for superintelligence.
Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.
- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.
So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.
If you learn AI Infrastructure Engineering, you will never be unemployed again.
Stage 1: Systems & GPU Foundations
- Learn: C++, Rust, CUDA memory hierarchy, PCIe vs NVLink, thread blocks and warp scheduling.
- Practice: Write a custom CUDA kernel for matrix multiplication from scratch and profile it against cuBLAS.
- Why: You cannot optimize what you do not understand at the silicon level.
Stage 2: Transformer Inference Physics
- Learn: Prefill vs Decode phases, arithmetic intensity, memory bandwidth bottlenecks, KV-cache math.
- Practice: Profile a HuggingFace model with Nsight Systems to identify the exact memory bottleneck during token generation.
- Why: LLM inference is almost always memory-bound, not compute-bound.
Stage 3: Modern Serving Engines & Batching
- Learn: vLLM, SGLang, PagedAttention, continuous batching, chunked prefill.
- Practice: Deploy a 70B model and tune chunk sizes to maximize throughput without starving decode requests during long-context prefills.
- Why: Naive batching leaves 60% of GPU VRAM wasted. PagedAttention fixes this.
Stage 4: KV-Cache & Memory Optimization
- Learn: Prefix caching, KV quantization, CPU offloading, multi-turn cache reuse.
- Practice: Build a routing proxy that directs requests with identical system prompts to the same replica to share KV blocks.
- Why: Reused cache is free speed. It can slash Time-To-First-Token (TTFT) by 80%.
Stage 5: Quantization & Compression
- Learn: FP8, INT4, AWQ, GPTQ, Sparsity, TensorRT-LLM calibration.
- Practice: Serve a model in FP8 vs FP16 and benchmark the exact perplexity drop against the latency and VRAM gains.
- Why: Quantization is the only way to fit frontier models on edge GPUs and protect your margins.
Stage 6: Kernel-Level Engineering
- Learn: Triton, FlashAttention, CUDA graphs, operator fusion.
- Practice: Write a fused Triton kernel for RMSNorm or Softmax and benchmark it against the native PyTorch implementation.
- Why: Python overhead kills inference. Fused kernels save milliseconds that compound at scale.
Stage 7: Distributed Inference & Parallelism
- Learn: Tensor parallelism, Pipeline parallelism, Expert parallelism (for MoE models).
- Practice: Shard a 405B parameter model across 8 nodes and measure the communication overhead between GPUs.
- Why: Single-GPU inference is dead for frontier models. You must master the shard.
Stage 8: Speculative Decoding
- Learn: Draft-target models, Medusa heads, acceptance rates, n-gram drafting.
- Practice: Build a pipeline where a local 3B model drafts tokens for a cloud 70B model to verify in parallel.
- Why: 2x decode speed at zero quality cost is the closest thing to a free lunch in inference.
Stage 9: Multi-Node & Hardware Interconnects
- Learn: NCCL, RDMA, InfiniBand, NVLink, Disaggregated Prefill/Decode architectures.
- Practice: Set up a multi-node cluster and profile the network latency of tensor parallelism across nodes vs within a single node.
- Why: Network latency is the new GPU bottleneck. Disaggregating prefill and decode is the 2026 meta.
Stage 10: Cluster Orchestration & GPU Scheduling
- Learn: Kubernetes GPU operators, Ray, Slurm, MIG (Multi-Instance GPU) partitioning, KEDA.
- Practice: Build a queue-based autoscaler that spins up spot GPUs based on pending inference requests and drains them when empty.
- Why: Idle H100s burn $3+/hr. FinOps and scheduling are now core infra responsibilities.
Stage 11: AI Gateways, Routing & Observability
- Learn: TTFT/ITL SLOs, semantic routing, DCGM metrics, OpenTelemetry for LLMs.
- Practice: Build a gateway that routes simple queries to a quantized local model and complex reasoning to a frontier API based on prompt complexity.
- Why: Routing protects your margins and DCGM metrics tell you when your GPUs are silently throttling.
Stage 12: Public Benchmarks & Teardowns
- Learn: Reproducible methodology, latency/throughput Pareto curves, cost-per-token analysis.
- Practice: Publish a teardown comparing vLLM vs SGLang vs TensorRT-LLM on your specific hardware with full configs.
- Why: Public proof of hardware mastery gets you hired instantly by top AI labs.
Wrappers are a commodity. AI Infrastructure is the physics of scale.
The modern AI Infra Engineer builds the silicon nervous system for global intelligence.
Bookmark and Repost!
My biggest takeaway from SF Tech Week: the center of gravity has already moved.
Last year everyone was talking about agents.
This year, a lot of the interesting conversations were about everything agents need once they actually start working: routing, search, data, evals, observability, latency, inference cost, GPUs, deployment, reliability.
The question used to be:
Can an agent do this?
Now it increasingly feels like:
How do we make a million agents do this reliably, cheaply, and controllably?
That shift matters.
Companies like Tavily are a good example. Most end users may never know the name, but a huge number of agent workflows need search and web retrieval underneath. Once agents become real products, the boring-looking infrastructure around them becomes extremely valuable.
Another thing Tech Week messes with is your perception of startup outcomes.
You meet founder after founder who raised, got acquired, or somehow went from starting a company to becoming a VP / GM / Head of X at a much larger company in a few years.
After hearing enough of these stories, your brain starts treating this as normal.
It obviously isn't.
Tech Week is basically the Instagram feed of the startup distribution. You are meeting people from the far right tail. The companies that ran out of money, never found PMF, or quietly dissolved are not usually on stage telling you their story.
And even we got acquired can mean very different things: a great strategic outcome, an acquihire, investors getting some money back, or simply a soft landing.
Startups don't just increase expected upside. They massively increase variance.
One other thing I kept hearing was FDE: Forward Deployed Engineer.
The role is interesting because it blurs boundaries that used to feel much cleaner.
Engineer builds.
Sales sells.
PM translates.
But more and more, the valuable person seems to be someone who can talk to the customer, understand the business problem, talk to engineers, turn ambiguity into something executable, and actually ship it.
Engineer + Product + Consultant + Sales.
I increasingly think this is less a job title than a survival skill.
AI is already crawling into every industry and every workflow. So maybe the interesting question is no longer what AI can do.
It is:
What workflow has not been rebuilt yet?
And after spending a few days around all of this, I came away with two contradictory feelings:
How is everyone here doing something this ambitious?
And:
Maybe I should be doing more too.
The Bay Area has never had a shortage of ideas or smart people.
The scarce things are judgment about what is worth doing, the willingness to stay with it for years, and the ability to get other people to care.