The Sequence Radar - Issue #958: Last Week in AI: Cheap Intelligence, Expensive Ambitions
Next Week in The Sequence:
We continue our series about recursive self-improvement.
We dive into DeepSeek’s inference stack, Google’s new agentic VM platform and more.
We write about the AI credit crisis.
And we will have awesome installment about robotics startups
Subscribe and don’t miss out:
TheSequence is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.
📝 Editorial: Last Week in AI: Cheap Intelligence, Expensive Ambitions
Artificial intelligence is getting more efficient, which apparently means everyone needs to spend more money on it. This week, DeepSeek released inference software, NVIDIA offered faster AI factory construction, Meta and SpaceX reported enormous investments, and Legora collected another $550 million to help lawyers work. The budget meeting must be going beautifully.
DeepSeek’s contribution was DeepInfer, the framework behind its production inference services. The open release supports DeepSeek-V4.1-Flash on NVIDIA and Huawei hardware, using techniques such as drafting tokens before verifying them and overlapping scheduling with computation. It provides an inference core; operators still need caching and load balancing for a complete service. The economic ambition is straightforward: squeeze more output from expensive hardware. If serving each request becomes cheaper, developers can afford agents that make many more requests. Efficiency can sell GPUs. Intelligence is impressive. Intelligence with a better gross margin gets a spreadsheet.
At GTC Berlin, NVIDIA introduced DSX POD, a design for modular Vera Rubin AI factories. NVIDIA says deployment at prepared sites can fall from six months to as little as 14 days, with partner deliveries expected in early 2027. The prepared-site qualification matters: permits, electricity and cooling have not suddenly become downloadable software. Its new OpenShell Reference Architecture adds infrastructure for governing and securing agents. The sales pitch extends from producing intelligence to installing the factory and supervising the employees.
Meta supplied evidence that AI can improve a business that already makes money. Quarterly revenue rose 26% to $64.6 billion; management credited AI with improving advertising through better engagement and recommendations. Yet quarterly capital spending reached $34.9 billion, and free cash flow was negative $2.1 billion. The existing business helps finance the ambition. The ambition is getting remarkably good at absorbing the financing.
SpaceX offered an even stranger corporate identity. Its AI division, now called SpaceXSI, generated $7.56 billion in quarterly revenue, exceeding Starlink’s $5.25 billion. Cloud services, the Cursor acquisition and Grok contributed to growth. The AI segment reported $431 million in operating income and $16.44 billion in capital spending. You can run a rocket company and discover that the truly expensive expansion project is helping people run software.
Then there is Legora, which raised $550 million at an $8.8 billion valuation and says annual recurring revenue now exceeds $200 million. Its legal software points toward a different pricing logic: customers care about the cost of getting work done. A token may be getting cheaper; a lawyer’s afternoon remains a valuable reference price. The investment thesis is that owning the workflow lets you keep some of the difference. Whether customers or software vendors keep those savings will be an interesting negotiation, presumably involving lawyers.
The wager connecting these announcements is that cheaper inference expands demand enough to justify more new infrastructure, while specialized applications capture part of the value. That is plausible. It also requires distinguishing revenue growth from attractive returns on the capital needed to produce it. Everyone can have a growth story. They still need to agree on whose margin it is.
🔎 AI Research
Self-Evolving Coding Agents via Harness-Policy Optimization
AI Lab: Microsoft Research, University of Virginia, Carnegie Mellon University
Summary: This paper introduces Self-HPO, a recursive self-improvement framework for coding agents that alternates between optimizing the harness and the policy model, with the evolving policy proposing its own harness changes after analyzing its failures. Each round picks a harness, trains the policy with reinforcement learning under it, and feeds the hard tasks surfaced in training into the next harness search. Across three coding benchmarks it lifts Qwen3.5-35B-A3B’s average resolve rate from 57.8% to 64.8%, beats harness-only and policy-only baselines, and shows a joint gain larger than the sum of the two individual gains.
Cola2: Co-Adaptive Hierarchical Language Modeling with Continuous Semantic Compression and Balanced Discrete Generation
AI Lab: ByteDance Seed, The University of Hong Kong, Peking University, Renmin University of China
Summary: This paper introduces Cola2, a hierarchical language model that separates continuous semantic modeling from discrete text generation by compressing text into dynamically segmented semantic spans, generating those latents with a flow-matching prior, and realizing tokens with a block-causal masked diffusion decoder, with the encoder and decoder co-adapted through token-prediction gradients. With a Qwen3-4B-Base-initialized decoder, an unconverged checkpoint trained on 346 billion source tokens reaches a 60.10 average across eight tasks and surpasses Llama-3-8B on HellaSwag and StoryCloze despite a reported pretraining budget more than 43 times larger.
Foundation4D: Towards a unified perception and reconstruction model for autonomous vehicles
AI Lab: NVIDIA
Summary: This paper introduces Foundation4D, a unified feed-forward model that estimates geometry, 3D objects, semantic occupancy, and Gaussian scene reconstructions from multi-camera driving video, using a shared backbone with factorized spatiotemporal attention across cameras and time and a video-level pipeline that applies pose graph optimization to refine trajectories and assemble object tracks. With benchmark-specific fine-tuning it reaches the lowest depth error on Waymo and nuScenes, 60.9 mAP and 66.5 NDS on nuScenes detection, and 27.4 IoU and 18.1 mIoU on OpenOccupancy, and the authors also release PAI-Foundation4D, a benchmark with dense depth and occupancy labels for 1,253 driving clips.
SplatLab: Autonomous Scientific Discovery for 3DGS Research
AI Lab: University of California, San Diego
Summary: This paper introduces SplatLab, an autonomous research system for 3D Gaussian Splatting that reproduces a published baseline, proposes an extension, runs training and ablations on Mip-NeRF 360 and LLFF, and writes the manuscript with every reported number bound to a logged training run. It evolves candidates through a cascade-weighted Elo tournament across eight islands with a standing falsifier that subtracts Elo when a candidate fails stress tests, and its manuscripts beat their reproduced targets by 0.18 to 3.78 dB PSNR, including 3.78 dB on astrophotographic nebula rendering, in 24 to 48 hours per run.
Agent World Model in the Wild: Computer-Use Agent Environment Synthesis
AI Lab: Microsoft Research, UNC-Chapel Hill, Georgia Tech, Northeastern University
Summary: This paper introduces WildAWM, a pipeline that uses only open-source models to synthesize persona-based personal computers where websites, MCP services, and local files share one evolving state, then generates cross-interface computer-use tasks that must pass five execution-based quality gates. The resulting 341 long-horizon environment-task pairs can take over 100 steps and GPT-6-Astra solves only about 60% of them, and Qwen3.6-35B-A3B and Qwen3.5-4B trained exclusively in these worlds improve on all eight out-of-distribution web, claw, and computer-use benchmarks by 10.58 and 8.21 points on average.
Retrieval Augmented Generation Compositionality
AI Lab: IBM, Weizmann Institute of Science
Summary: This paper introduces LASAGNA, an open-source framework for controlled RAG evaluation with eight swappable pipeline stages and 55 built-in components, and uses it to test 450 configurations across eight benchmarks and corpora from 5K to 200K documents. It finds that retrieval gains do not add up, since three techniques predicted to lift NDCG@10 by 47% deliver only 17% together, while faithfulness gains combine nearly additively at +28% predicted and +32% realized, and that the best pipeline on small corpora trails by 13.7% at 200K documents, with the ranking flip visible from 25K.
Advancing Coding in Low-Resource Programming Languages
AI Lab: OpenAI
Summary: OpenAI introduces two evaluations for coding in low-resource languages such as COBOL, OCaml, Fortran, and Visual Basic: MultiLangBench, which asks a model to build a SQLite-compatible database engine from scratch across 20 languages and is being released on GitHub, and a COBOL migration benchmark that requires byte-for-byte output parity after porting programs to C++, Java, Kotlin, Python, or Rust. GPT-6 Astra reaches an 87.4% average pass rate on MultiLangBench, up from 26.8% for GPT-5.6 Sol, with COBOL rising from 10.9% to 72.1%, while GPT-6.1 Sol hits 86.6% on COBOL migration at low reasoning effort, up from 5.4%, and 97.9% at higher effort.
LLMs Are Making Formal Software Verification Feasible at a Much Larger Scale
AI Lab: Anthropic (Frontier Red Team), in collaboration with Theorem
Summary: In collaboration with Theorem, Anthropic used Claude to turn developer fuzz tests into formal specs over compiled RISC-V binaries and then write the proofs, showing that 239 entry points across 123 open-source OSS-Fuzz libraries never crash on any input under 64 MB, with proofs totaling about 180 million lines of Rocq that an external proof checker verifies. Along the way Claude patched hundreds of bugs that fuzzers were unlikely to find, including an out-of-bounds write that needed a roughly 20 MB input, and the team found that Opus 5.5 needs about 40% fewer tokens than Opus 5 on a 36-task proof benchmark and that agents sharing proofs used 17% fewer tokens, while cautioning that “formally verified” does not mean bug-free since the proofs assume single-threaded static RISC-V binaries on Linux.
🤖 AI Tech Releases
DeepSeek DeepInfer
DeepSeek open sourced DeepInfer, the in-house inference framework behind its production DeepSeek-V4.1-Flash serving, releasing it under Apache 2.0 with optimized kernels, MoE expert parallelism, speculative decoding, and deterministic sampling for NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs.
NVIDIA Run:ai Open Source
NVIDIA published Run:ai as open source, releasing the full GPU and AI workload orchestration platform for Kubernetes (control plane, cluster services, installer, CLI, and MCP server) with fair-share scheduling, fractional GPU allocation, and multi-cluster management.
Google Cloud Agent Compute Engine
Google Cloud introduced Agent Compute Engine, the hardware-isolated microVM service that powers Gemini’s agents, coming to preview in the coming months alongside the open-source Agent Substrate runtime for GKE and expanded Agent Gateway governance for agents running across clouds, on-prem, and endpoints.
NVIDIA Nemotron 3.5 Super VL and Bolt
NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter (12B active) multimodal model for always-on enterprise agents that it says is up to 5x faster than comparable open models, alongside two edge models, Bolt 74B-A9B 24GB for local agents on RTX and DGX Spark and Bolt E4B 4GB for fast conversational agents, which arrived October 23, all with open weights, datasets, and training recipes.
📡 10 AI News You Need to Know About
Legora Series E: Legora raised a $550 million Series E at an $8.8 billion valuation, led by Atreides Management with Fidelity, Capital Group, and J.P. Morgan joining, as the legal AI company passes $200 million in ARR and expands its agentic Legora aOS platform across North America, Europe, and Asia-Pacific.
Warp Sciences Series A: Former OpenAI science lead Kevin Weil launched Warp Sciences with a $215 million Series A led by Lightspeed Venture Partners, valuing the company at $865 million according to Bloomberg, to build robotic, fully automated labs that generate training data for AI models and speed up drug development in the US.
OpenAI Funding Round: OpenAI is reportedly nearing a round of up to $30 billion at a $1.25 trillion pre-money valuation led by Coatue and investors from the UAE, according to CNBC, which also reports a possible compute-capacity deal with the UAE and a target of securing signatures by the end of next week.
DeepSeek Funding: DeepSeek has raised about 100 billion yuan, roughly $15 billion, according to Bloomberg, in a round backed by Tencent and CATL among others after doubling its original 50 billion yuan target, as the Chinese lab prepares for a potential IPO.
Recursive Flint: Recursive, the self-improving AI startup led by Richard Socher, released Flint, an open-weight 40B-parameter mixture-of-experts coding model that activates 2.5B parameters per token and runs locally on laptops, built in under four months with the company’s automated research system, the Eureka Machine.
OpenAI Acquires Orchestra Bio: OpenAI acquired Orchestra Bio, a platform whose AI agents help biopharma teams coordinate drug-development programs, and will fold its team and technology into Rosalind, OpenAI’s models and tools for life sciences research; terms were not disclosed.
Nscale and Anthropic: Nscale is in talks to sell Anthropic a second roughly $45 billion block of computing power, covering a second 460-megawatt building at its Monarch campus in West Virginia built on Nvidia Vera Rubin chips, according to Bloomberg, which says a deal could be finalized within days but details could still change.
Tenstorrent Series E: AI chip company Tenstorrent closed a $1.5 billion Series E at a $6.2 billion post-money valuation, led by Trace Capital and Innovation X, to deploy its Galaxy servers at scale for open-weight LLMs and video generation and to expand the AI and CPU IP it licenses to chip designers.
Meta Q3 Earnings: Meta reported third-quarter revenue of $64.6 billion, up 26% year over year, with capital expenditures of $34.9 billion and negative free cash flow of $2.1 billion as it keeps building AI infrastructure, narrowed its 2026 capex outlook to $135 billion to $145 billion, and said its board approved a new class of non-voting Class C common stock, effective January 19, 2027, that it can use to raise capital, pay employees, and fund acquisitions for its AI buildout without diluting the voting power of existing shareholders.
SpaceX Super Intelligence: SpaceX’s AI business, now reported as its Super Intelligence segment, generated $7.56 billion in third-quarter revenue, up 830% year over year, and turned a $431 million operating profit on cloud services agreements, the Cursor acquisition, and Grok adoption, after spending $16.4 billion on AI capex to reach 1.7 gigawatts of compute, with Elon Musk telling investors on the earnings call that SpaceX expects 2.5 to 3 gigawatts online by year-end and is fairly confident it can add 10 more gigawatts in 2027.