The Sequence Radar #755: Last Week in AI: Worlds Built, Models Refined, and Legends Move On

World models dominated last week in AI but that wasn't all.

Created Usiung GPT-5

Next Week in The Sequence:

Our series about synthetic data generation continues with an intro to generative synthesis. In the AI of the week, we are going to dive into DeepMind’s new SIMA world model. Our opinion section will dive into the possibilities and challenges of world models.

Subscribe and don’t miss out:

📝 Editorial: Last Week in AI: Worlds Built, Models Refined, and Legends Move On

Pretty loaded week in AI across several trends: world-models, large language models, and multilingual speech—are starting to converge into a more coherent picture of where the field is heading. We also have major fundraise events and some surprising departures.

On the world-model side, Marble from World Labs illustrates how far things have moved beyond static perception. Rather than just taking in images or video and labeling what’s there, Marble is designed to reconstruct persistent 3D environments from multimodal input and maintain coherent state over time. The model can assemble scenes, objects, and surfaces into a navigable representation that agents can inhabit and interact with. For practitioners, this suggests a different kind of data and tooling stack: pipelines for 3D assets and scene graphs, training loops that care about continuity and physical plausibility, and evaluation focused on how well an environment supports downstream tasks rather than just reconstruction quality.

World Labs itself is a good illustration of how individual researchers shape these trajectories. As a co-founder, Fei-Fei Li brings a track record that includes ImageNet, large-scale visual recognition benchmarks, and a sustained push toward data-centric AI. Her work has repeatedly redefined how the field thinks about supervision, representation, and the role of high-quality datasets. In many ways, Marble can be seen as an extension of that vision: moving from static labeled images to rich, interactive worlds as the substrate on which intelligent behavior is learned. Her broader impact—through research, open datasets, and institution-building—is part of why world-models are emerging as a serious, well-founded direction rather than just a speculative trend.

DeepMind’s SIMA-style agents push the same idea from a different angle: controlling agents that operate inside complex virtual worlds. Instead of treating a world-model as a passive representation, SIMA emphasizes interactive behavior—navigating spaces, manipulating objects, following natural-language instructions, and learning from feedback within 3D environments. This brings together perception, action, and language into a single loop. If Marble is about building the stage, SIMA is about training the actors. For the ecosystem, that means more attention to interfaces between foundation models and engines, from standardized action spaces to APIs for logging trajectories and rewards at scale.

Before turning to language models, another notable development this week came from the developer-tooling ecosystem: Cursor announced a $2.3 billion funding round, lifting its valuation to $29.3 billion. Cursor has quickly become one of the most widely adopted AI coding assistants because of its tight integration between the editor, the model, and the developer workflow. The new funding round signals that AI-assisted software engineering is moving from novelty to necessity, with tools increasingly expected to reason over entire codebases, manage iterative changes, and support autonomous refactoring. For engineering teams, this validates a broader shift: the future of IDEs won’t be defined by plugins that bolt AI onto legacy workflows, but by environments designed from the ground up around collaboration between humans and models.

In parallel, large language models are entering a refinement phase. The release of GPT-5.1 is less about a shocking capabilities jump and more about making high-end models more controllable, more consistent, and more aligned with real workflows. The new variants emphasize better instruction following, adaptive reasoning—deciding when to think briefly versus in more depth—and richer persona control so that the model can respond in styles tuned to different products and audiences. The signal here is clear: for many real-world applications, the bottleneck is no longer “can the model do it?” but “can the model do it in a way that is predictable, steerable, and pleasant to use?”

Speech and language coverage also took a meaningful step forward with the release of a large-scale omnilingual ASR system capable of transcribing an enormous number of languages, including many that have historically been neglected in machine learning benchmarks. By bringing hundreds of low-resource languages into a reasonable error range and offering an open ecosystem around the models and data, this kind of system materially lowers the barrier for building voice interfaces, accessibility tools, dubbing and captioning pipelines, and localized products for communities that have never really been served by mainstream AI.

Overlaying all of this is a major human signal: Yann LeCun’s decision to leave Meta marks the end of one of the most influential tenures in industrial AI research. As Chief AI Scientist, he helped shape Meta’s long-term bet on self-supervised learning, championed the development and deployment of large-scale computer vision and recommendation systems, and pushed for a more open, research-centric culture through FAIR. His work inside Meta often provided a counterweight to the industry’s narrow focus on language-only models, emphasizing energy-efficient architectures, representation learning, and long-horizon autonomy. His departure closes a chapter in which Meta played a central role in advancing deep learning at scale, and it raises important questions about how the company will define its research identity going forward.

Quite a week! Let’s dive into the details.

🔎 AI Research

SIMA 2 – An agent that plays, reasons, and learns with you in virtual 3D worlds

AI Lab: Google DeepMind / Google DeepMind Games & Agents team

Summary: SIMA 2 is a Gemini-powered generalist agent that operates in commercial and AI-generated 3D games by “seeing” the screen and controlling a virtual keyboard and mouse, following natural-language, sketch, and emoji instructions. It goes beyond the original SIMA by setting its own goals, explaining its plans, generalising skills across games, and continuing to improve through self-play in both human-made and Genie-generated worlds.

Understanding neural networks through sparse circuits

AI Lab: OpenAI

Summary: OpenAI trains weight-sparse transformer models whose internal computations decompose into small, disentangled “circuits,” making it possible to isolate the subnetworks responsible for specific behaviors instead of reverse-engineering a dense tangle of weights. By scaling these sparse models, they show you can maintain strong capabilities while making internal mechanisms more transparent, pointing to a path for safer, more interpretable AI systems. o

LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics

AI Lab: Meta FAIR; New York University (NYU); Brown University

Summary: Introduces LeJEPA, a JEPA objective grounded in a proof that isotropic Gaussian embeddings minimize worst-case downstream risk, enforcing this via SIGReg—a sketched Epps–Pulley characteristic-function test over random 1-D projections—thus removing stop-gradients, teacher–student schemes, and other brittle heuristics with a single trade-off λ and linear time/memory. Across 60+ architectures (scaling to ~1.8B-parameter ViT-g) it trains stably, outperforms frontier transfers in in-domain settings, and its training loss closely tracks linear-probe accuracy, enabling label-free model selection.

Llama-Embed-Nemotron-8B: A Universal Text Embedding Model for Multilingual and Cross-Lingual Tasks

AI Lab: NVIDIA

Summary: Presents an open-weights 8B embedding model that ranks #1 on the MMTEB (multilingual) leaderboard via a bi-directional Llama-3.1 encoder, diverse synthetic+non-synthetic training, and model-merging. It delivers strong retrieval/STS/classification across 250+ languages and details ablations on loss design, SDG models, and merging.

MMCTAgent: Multi-modal Critical Thinking Agent Framework for Complex Visual Reasoning

AI Lab: Microsoft Research India

Summary: Introduces a planning-and-tools framework plus a vision-based critic that verifies answers against automatically generated criteria, boosting image/video QA. The agent outperforms strong MLLMs and tool pipelines on MMVET/MMMU and long-form video (EgoSchema), with the critic adding further gains.

Black-Box On-Policy Distillation of Large Language Models

AI Lab: Microsoft Research

Summary: Introduces GAD (Generative Adversarial Distillation), a black-box, on-policy distillation framework where a student LLM learns via a discriminator that compares its outputs to a teacher’s, forming an adversarial minimax game. Experiments show GAD significantly outperforms standard sequence-level distillation, achieving strong generalization and enabling small open-source models to approach GPT-5-Chat performance.

🤖 AI Tech Releases

Marble

World Labs launched Marble, its first text to 3D World model and its formidable.

GPT 5.1

OpenAI launched GPT-5.1, two new models that optimize the conversational experiences in ChatGPT.

Omnilingual ASR

Meta released a series of automatic speech recognition models that enable speech intelligence capabilities for over 1600 languages.

ERNIE-4.5-VL-28B-A3B-Thinking

Baidu released RNIE-4.5-VL-28B-A3B-Thinking, a new moultimodal model optimized for reasoning workflows.

JAX-Privacy 1.0

Google open sourced JAX-Privacy, a library for differentially private AI.

📡AI Radar

• Cursor – AI code editor rockets to mega-round ( $2.3B Series D at $29.3B valuation )
  AI coding platform Cursor closed a $2.3B Series D at a $29.3B post-money valuation, after 100× year-to-date growth in enterprise revenue, to scale its Composer model and deepen product development for professional developers.

• Yann LeCun – Meta’s chief AI scientist prepares a new venture
  Turing Award–winner and FAIR founder Yann LeCun is reportedly planning to leave Meta to start his own AI startup, and is already in early fundraising talks; his departure follows Meta’s internal AI reorganization under the new Superintelligence Labs unit.

• Milestone – measuring whether GenAI actually pays off ( $10M seed )
  Milestone raised a $10M seed round led by Heavybit and Hanaco to help engineering leaders track how AI coding tools affect productivity and code quality, turning GenAI usage into measurable ROI dashboards.

• Bindwell – teen founders apply AI to pesticide discovery ( $6M seed )
  Bindwell, founded by two teenage engineers, raised a $6M seed round to use custom AI models (borrowed from drug discovery) to design safer, more effective pesticide molecules and compress lab cycles from days to seconds.

• WisdomAI – AI “data analyst” for enterprises ( $50M Series A )
  WisdomAI, which builds conversational AI “data analysts” on top of an enterprise context layer, raised a $50M Series A led by Kleiner Perkins with Nvidia’s NVentures and others, bringing total funding to about $73M to scale its proactive analytics platform.

• Uare.ai (formerly Eternos) – from digital immortality to personal AI clones ( $10.3M seed )
  Eternos has rebranded as Uare.ai and raised $10.3M seed funding to build “Individual AIs”—personal models trained on a person’s stories, values, and voice that can act as a scalable digital twin for creators and professionals.

• Kaltura + eSelf.ai – AI avatar tech folded into enterprise video ( ~$27M M&A )
  Kaltura signed a definitive agreement to acquire eSelf.ai, a multimodal AI lab for interactive avatars, in a deal totaling roughly $27M in cash and stock to embed expressive, talking agents across its video and learning products.

• 1mind – AI “superhuman” for sales teams ( $30M Series A )
  1mind, founded by 6sense’s Amanda Kahlow, raised a $30M Series A led by Battery Ventures (total funding ~$40M) to scale “Mindy,” an AI sales agent that handles inbound leads, demos, and onboarding using a blend of deterministic logic and LLMs.

• Scribe – workflow AI and “where to use AI” map ( $75M Series C at $1.3B )
  Scribe, known for auto-documenting workflows, raised a $75M Series C led by StepStone at a $1.3B valuation to launch Scribe Optimize, which analyzes how work gets done across apps and flags where automation and AI will deliver real ROI.

• Inception – diffusion LLMs for code and text ( $50M round )
  Inception, led by Stanford professor Stefano Ermon, raised $50M from Menlo Ventures, Microsoft’s M12, Snowflake, Databricks and others to develop diffusion-based large language models (“dLLMs”) that promise faster, cheaper inference for code, text, and real-time applications.

• Wabi – “YouTube of apps” from the Replika founder ( $20M pre-seed )
  Replika founder Eugenia Kuyda’s new startup Wabi raised a $20M pre-seed round to build a social platform where users describe an idea in natural language and instantly get a shareable mini-app that others can remix, like, and reuse.

• Firmus – green AI data centers in Australia ( ~US$327M / AU$500M raise )
  Australian neocloud firm Firmus secured about AU$500M (US$327M) in equity to expand Project Southgate, a 1.6GW “green AI factory” data-center initiative powered largely by renewables across Tasmania and mainland Australia. Link: Firmus site and Southgate overview. (Firmus)

• Netic – AI revenue engine for trades and home services
  Netic, which builds autonomous AI agents to answer calls, book jobs, and run revenue operations for plumbers, roofers, HVAC and other “backbone” service industries, raised a $23M Series B led by Founders Fund to extend its earlier $20M seed and Series A stack.

• FMC (Ferroelectric Memory) – low-power memory for AI era
  Dresden-based FMC raised €77M in Series C equity plus €23M in subsidies (total €100M) to commercialize its DRAM+ and 3D-CACHE+ ferroelectric memory chips, aiming to cut power use and boost performance for AI and data-center workloads.

• Neysa – AI cloud platform draws interest from mega-investors
  Indian AI cloud startup Neysa Networks, which offers a full-stack GPU-centric “Velocis” AI cloud system, is reportedly in talks to sell stakes to Blackstone and SoftBank at a sub-$300M valuation, potentially giving it major capital and distribution for its GPUaaS and AI PaaS services.

• Tencent – earnings beat under “measured” AI strategy
  Tencent’s Q3 2025 results showed revenue up ~15% year-on-year to ~RMB193B and net profit up about 19–20%, with management explicitly crediting AI investments for better ad targeting, gaming engagement, and internal efficiency; the company continues to integrate its Hunyuan foundation model across products while keeping AI capex relatively disciplined.