It's Lunar New Year, so here's my reflection for 2025: (1) Building more practically applied AI services and agentic engineering I accepted that the AI era already started, and my job now is not to try to get offended by its capabilities, but to learn how to master the technology. I switched entirely from VSCode to @cursor_ai and then Codex, Gemini, and Claude. I was aware of the AI advancement back in 2022, but only using @OpenAI for a daily chatbot. Even the AI-applied services I built before only stop at tuning prompt techniques. So I dived deep into RAG, inference, and advanced frameworks like @LangChain_JS and @crewAIInc to build apps like @GrowlySuite and @speedgrowth_ai (2) Seize all the opportunities, and life will reward you I’ve realized that life doesn't just hand out participation trophies; it rewards the intensity with which you seize the opportunities in front of you. This year was a testament to the fact that when you lean into the chaos of a changing industry and put in the raw effort, the universe tends to meet you halfway. My team at @GrowlySuite got a chance to live in @ns for a month, and it was a life-changing experience, seeing people building in the frontier with innovative ideas. Or I still remember the first time I spoke in a big conference like @gmvn_official to represent @Polkadot. At the end, it’s about being positioned to say "yes" when the right door opens, and having the stamina to walk through it. (3) Health is still the strongest weapon you need to sharpen every single day In the midst of building and scaling, I’ve been reminded that health isn't a secondary goal—it is the strongest weapon in my arsenal. You can have the most advanced AI agents in the world, but if the "human in the loop" is compromised, the system fails. (4) Loss is a part of life 2025 also carried the heavy reminder that loss is an inextricable part of the human experience. Whether it is the loss of people, opportunities, or past versions of ourselves, I’ve had to learn how to carry that weight without letting it anchor me in the past. It is a quiet, difficult teacher, but it has added a layer of depth to my work and my relationships that wasn't there before. (5) This all shall pass "This all shall pass" has become my mantra for the upcoming year. I am still learning how to truly accustom myself to the fact that time is fleeting—that the highs are temporary and the lows are, too. This perspective brings a strange kind of peace; it forces me to stay present in the build and grateful for the moment, knowing that the only constant is change.

Liquid Inference now supports images and video! Video inference costs aren’t optimized to the same level as text, and a fair auction brings iterative AI content creation in line. Access via our API or check out the compare tool in our web playground.

The Liquid Inference router network is growing quickly. We now have 28 verified companies serving >1000 models, offering capacity dynamically and updating quotes in real-time. Average time for providers to onboard and integrate models is 3 hours, full verification within a day.

The experience is so seamless that you only need a few clicks to get ready.

Stats from Liquid Inference’s first few days live: • Cheaper token prices than 98% of OpenRouter traffic • ~2B tokens served • 24 verified providers onboarded to the network • 88% cache hit-rate • 26 harnesses/IDEs supported Buckle up, many new features launching this week.

Today we’re launching Liquid Inference, an OpenAI/Anthropic-compatible router where inference providers compete on price for every prompt. Buyers route across 700+ open- and closed-weight models at the lowest available marginal cost, providers onboard instantly to start serving models.

Introducing Liquid Inference by Architect, a new LLM model router where inference providers compete dynamically to offer lowest-cost tokens. Providers onboard in minutes, not weeks. Buyers have full price visibility and receive 20% of referred fees in free inference. More info: Liquid Inference can be used in place of any OpenAI or Anthropic API-compatible backend. Behind it, our team has leveraged our experience trading and building financial exchanges to create real-time two-sided price discovery for inference. Benefits for buyers and providers: INFERENCE BUYERS • Fully compatible with agentic coding tools: Claude Code, Codex, OpenCode, Cursor, Pi, Cline, and many more. Full multi-modal support. • Hundreds of open- and closed-weight models. Create routing rule presets or use our Auto routing algorithms. • Providers compete dynamically on price for every prompt, offering the lowest marginal cost. • Sign up for free with email, first 500 users get $20 of free inference. • Receive 20% of referred fees as free inference, 10% for second-level referrals. INFERENCE PROVIDERS • Onboard through the Liquid Inference app. We verify new providers in minutes, not weeks. • All prompts are OpenAI API standard. Simple REST/Websocket API to register models and quotes. • Update quotes dynamically based on your own costs. Monetize GPU capacity only when you want. • Standardize payouts via Stripe, full itemized records of all jobs performed.

I stayed up til 2am and spent $1,000 benchmarking Jev Router so you don't have to. Performance on DeepSWE was roughly the same as GPT-6 Astra on low. It costs slightly more, and it took almost 5x longer to run.

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…

GPT-6 Astra, trained on ~100K+ NVIDIA Grace Blackwell NVLink72. From ChatGPT to o1 to Astra in 4 years. AGI has arrived. Congratulations @OpenAI team. 400K GPUs coming online next.

Architect is hiring for three lateral roles: options trading quantitative researcher, compute capital market sales, and structured derivatives specialist. Details on what we’re building and the new team members we’re looking for Quantitative Researcher: We’re building: automated high-frequency options trading strategies on legacy and new(er) exchanges What you’ll work on: volatility modeling, ML for parameter optimization, order placement simulation, new venue exploration Stack: Rust, Python, AWS, ClickHouse What we’re looking for: field experience in derivatives trading and research, strong system design instincts, multiple approaches to options pricing theory, bias towards shipping Nice to have: familiarity with the latest ML toolkits and exchange microstructure Location: your preference Compute Capital Markets Sales: We’re building: the first US-regulated exchange purpose-built for compute futures, options, swaps, forwards, and capacity, plus AI supply chain commodities What you’ll work on: building relationships and facilitating trades with neoclouds, lenders, insurance providers, inference companies, datacenters, AI labs, interdealer brokers, retail brokers, hedge funds, and asset managers What we’re looking for: network of established and startup AI companies, fluency with compute capacity pricing/structures, translation skills between finance and AI, aversion to slop, commitment to competitive markets for compute Nice to have: experience in compute procurement at a hyperscaler, neocloud, or lab Location: SF Structured Derivatives Specialist: We’re building: specialized contracts for trading and hedging compute, compute derivatives, and a range of physical commodities supporting the US datacenter build-out What you’ll work on: modeling and structuring OTC products with Architect’s traders/technologists and AI counterparties for compute, memory, inference, and energy trading What we’re looking for: experience trading OTC derivatives contracts at banks or trading firms, sell-side institutional network, enthusiasm for working at startups Nice to have: relevant FINRA/NFA certifications or passion for test-taking Location: your preference

LIVE: Join us to watch today's spacewalk as @Astro_Jessica and @Soph_astro work on maintenance outside the International Space Station. nitter.siddz.com/i/broadcasts/1RKZzBVPg…

How have software engineering fundamentals changed with agentic coding? Here is our AI Engineering Skills map for software engineering fundamentals.

Tungsten is coming to Architect’s AX. Perpetuals, not cubes. Hedge metals supply chain risk at every stage of compute manufacturing: metals mining, chip and memory fabrication, gas generation, electricity consumption, GPU capacity procurement, and inference creation.

Hedge funds will eventually provide inference to model routers for free, or even pay to do so, in order to get a first look at prompt contents. Payment-for-order-flow in the age of compute.

We've raised $700M at a $21B valuation from Jane Street, Kleiner Perkins, Sequoia, A16Z, Peter Thiel, BCV, and Blackstone. We're also excited to share that we've shipped our first rack to Jane Street.

I’ve built HFT/exchange systems in three languages that sit at very different points on the safety/performance/ergonomics spectrum: OCaml, C++, and Rust. Observations and a practical comparison from production use: OCaml • Strong static typing and algebraic data types make illegal states almost impossible to represent. Protocol parsers, order book models, and risk checks are self-documented by types and function signatures. • The default garbage collector is a non-starter for predictable tail latencies. There are workarounds using disciplined zero-allocation techniques: custom core libraries, careful use of the [@ inline] and [@ unboxed] attributes, and a culture that treats any heap allocation in the hot loop as a bug. • Functional programming forces strategy writers to think in pure transformations of market data to fair values to orders. This style enables multiple forms of unit testing and makes offline model training pipelines easier to implement without major code changes. • The ecosystem for kernel-bypass networking and FPGA offload is thin; performant OCaml requires writing a fair amount of C interop or custom FFI. Compile times and binary size can also become painful at scale. • Engineering effort on type safety is a replacement for future engineering effort on debugging. We used to half-joke, “if it compiles, it works.” C++ • The most commonly used language in HFT because it offers maximum control. The writer can pinpoint every cache line, allocation, and branch. When the last microseconds matter, few languages can compete. • Industry standard for exchange connectivity, FPGA drivers, and shared-memory IPC. Virtually every low-latency NIC vendor’s API is written for it. • The cost is constant vigilance. Use-after-free, data races, and subtle ABI mismatches are permanent time sinks. Even with modern C++ features such as concepts, spans, and std::atomic memory orders, the cognitive load stays high. • Template metaprogramming offers some of the type-level guarantees of OCaml or Rust, but the error messages and compile times can be inscrutable. • Most of the “C++ is faster” advantage narrows when factoring in the engineering time spent hunting undefined behavior. Rust • Ownership and borrowing give the memory-safety guarantees of a garbage-collected language with the performance profile of C++. Zero-cost abstractions allow HFT system authors to write zero-allocation, lock-free, cache-aware code and the compiler will refuse to allow introduction of data races. • Excellent for the modern low-latency stack: DPDK/AF_XDP bindings, io_uring, shared-memory rings, and safe wrappers around exchange binary protocols. • The type system is expressive enough to encode many of the same invariants of OCaml, such as phantom types, newtypes, and state machines, without the garbage-collection tax. • The borrow checker still occasionally forces awkward redesigns of hot-path data structures, async/await is not yet as mature for ultra-low-latency work as plain threads/ busy-polling, and the ecosystem for certain HFT-specific libraries lags C++. • It takes time for the ownership model to click, after which mental bandwidth can be dedicated to strategy and latency. It’s much more difficult to build a high-frequency trading system that retains correctness and readability as it scales than it is to build one that’s fast. Good taste and developer culture can coalesce around a particular language, but the best teams possess these qualities independently, and can adapt freely and agilely to new technologies as they become available.

In a 2023 podcast, Dwarkesh asked me if HFT was zero-sum for society. I argued that extreme competition at the margins for nanosecond-level advantage might subsidize the creation of broadly beneficial new technology. How AI training/inference stacks have become a clear example of this Over the last ten years, quantitative trading firms built or invested in the fundamental components of distributed systems and high-performance computing. Many technologies now commonly associated with AI were created for HFT during this period or reached scale because of it. A non-exhaustive list of these technologies, from high-throughput networking to vectorized databases to precision timestamping: • Mellanox InfiniBand and ConnectX NICs: HFT colocation and cross-connect environments drove early large-scale deployment of RDMA-capable InfiniBand and the ConnectX series. These became the default high-bandwidth, low-latency fabric inside modern GPU pods for all-reduce, parameter servers, and expert parallelism. Mellanox was acquired by Nvidia in 2020. • Solarflare/OpenOnload (and later Xilinx/AMD equivalents): Kernel-bypass networking stacks and specialized low-latency Ethernet NICs let applications DMA packets straight into user space. These techniques were required for software-based trading systems to reach the microsecond range of tick-to-trade latency. Engineers for hyperscalers commonly use this technology today for distributed cloud training and inference networks. • RDMA and RoCE: Remote Direct Memory Access, refined under HFT firms’ nanosecond requirements, has become the transport that allows GPUs to exchange gradients and activations with microsecond-scale latency and minimal CPU involvement. • FPGA feed handlers and inline accelerators (Alveo, custom RTL on Xilinx/Intel): FPGAs are not the hardware solution of choice for AI, where GPUs and custom ASICs have dominated. However, HFT use of FPGAs for wire-speed market-data parsing, book building, and sub-microsecond decision logic created a pipeline of talent and verification processes that transferred into AI inference accelerators. • Vectorized/columnar analytics engines: Quant and HFT shops require extreme scan and aggregation performance over massive tick and order-book histories, using pioneering technologies like KDB/Q. Several firms I’ve worked at became large users and funders of ClickHouse, which has become essential to logging and offline data preparation for large model training. • Hardware timestamping and PTP: HFT firms’ need for exact packet ordering and latency measurement hardened the timing infrastructure that distributed AI training and multi-node inference rely on for correct collective operations and debugging. • Lock-free, zero-allocation, cache-aware software patterns: The performance culture and specific techniques developed for tick-to-trade loops became standard practice in high-performance CUDA kernels, inference servers, and communication libraries. These techniques include carefully tuned concurrent data structures, huge pages, CPU pinning, and avoidance of syscalls in the critical path. The natural followup question in 2023 was whether too much capital and talent had been or could be spent on HFT. The HFT industry is still relatively small, employing around 10,000 people and producing about as much yearly revenue as Nvidia’s and AMD’s R&D budgets. If an ancillary use case for HFT is a bounty program to fund the infrastructure of our next technological epoch, the costs still seem low.

The Compute Desk Nvidia B300 GPU-hour index is at its all-time high. As margins on training grow thinner, neoclouds are collectively shifting focus to inference. High-priced B300s are generating ROI by lowering the marginal cost per token generated through batch inference.