AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
Latent Space
Reporting and perspectives from Latent Space. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Claude Mods are JavaScript or TypeScript plugins that run inside Claude Code and hook into events such as prompts, tool calls and interface renders, letting users reshape the CLI and Desktop UI at runtime. Anthropic ships six built-in mods, and Claude Code includes a skill that lets Claude write new mods from plain-language requests.

DeepSeek open sourced DeepInfer, the production inference stack behind V4.1 Flash, publishing figures of 155 trillion tokens per day from one cluster at $38 revenue per accelerator-hour. The release arrived on day one of PyTorch Conf, and vLLM later reported decode throughput within 6% of DeepSeek's numbers on GB200 in a matched configuration.

DeepSeek released DeepInfer, the Apache 2.0 inference engine behind V4.1 Flash, along with production figures including 155 trillion tokens per day from one cluster and a 96% input-cache hit rate. The KV-cache management and load-balancing components are not included in the release.

Tenex, the applied AI firm that pays engineers by story points rather than hours, has grown to 75 people with roughly half its customers now Fortune 2000 companies, and its engineering lead CJ Hess described at AIE New York how the company keeps project context in Git so agents can work from it.

Tenex, now 75 people with half its customers in the Fortune 2000, still ties engineering pay to story points, but says the real difficulty is compensating work that never lands in a ticket — architecture, client context and deciding what to build.

SF Compute has begun moving GPU capacity onto its order book, with 5% of total capacity tradeable so far, as CEO Evan Conrad described at AI Engineer New York. The company now runs much of a cloud's operating stack itself so buyers receive usable compute rather than a cash payout.

SF Compute has moved about 5% of its total capacity onto its trading order book as it shifts from reselling clusters to operating cloud-grade infrastructure itself, CEO Evan Conrad said at AI Engineer New York's new Compute Trading track.

DeepSeek's V4.1 Flash 1017 is a 552B-parameter, 16B-active open-weight MoE model under an MIT license with a 1M-token context window, priced unchanged at $0.30/$1.20 per million tokens; Artificial Analysis scores it 45 on its Intelligence Index, tied for second among open-weight models.

DeepSeek's V4.1 Flash 1017, a 552B-parameter open-weight MoE model with a 1M-token context and MIT license, lands near the top of open-weight rankings at $0.30/$1.20 per million tokens, and MiniMax's M3.1 Flash and Inception's diffusion model Mercury 3 also shipped.

Meta released Muse Spark Open, an 816B-total, 42B-active open-weight MoE with 1M context that Artificial Analysis scores 39, the highest for a US open-weights model. Reflection Beam 1.0, a 501B-total, 23B-active MoE under Apache 2.0, scored 30.

Meta released Muse Spark Open, an 816B-parameter open-weight mixture-of-experts model with a 1M-token context window, scoring first among US open-weight models on Artificial Analysis' Intelligence Index but trailing GLM-5.3 and Kimi K3. Separately, Goodfire's Nick Cammarata argues that cases of polysemanticity in neural networks dissolve under further agentic analysis.

Goodfire's Nick Cammarata published an essay arguing that most neurons in models up to trillions of parameters are already legible, and that many famous polysemanticity examples were misread; his critics say it does not resolve their key uncertainties.

Modal crossed $1B in annualized revenue and raised a $750M Series D at a $15.75B valuation led by Accel, with usage up more than 15x year over year. Its Sandboxes product alone is at a $300M run rate.

Modal has crossed $1B in annualized run rate and raised a $750M Series D at a $15.75B valuation, led by Accel, with usage tripling since its Series C and growing more than 15x year-over-year.

At AIE New York, OpenAI's Joseph Kim described a "threshold of usefulness" where models become reliable enough to save analysts time, and predicted agents will handle investment idea generation and trading by 2027.

Leash Bio is releasing a preprint, blog post and open dataset for Minos, an AI playground where agents run verifiable lab-in-the-loop scientific reasoning simulations built on the company's DNA-encoded library data.

Leash Bio released Minos, a preprint, blog post, and open dataset built from its DNA-encoded library data, giving AI agents tens of thousands of lab-in-the-loop simulations for scientific reinforcement learning. The work follows its Kaggle competition and Hermes model.

Google's Gemini 4 Argon is now generally available across the paid Gemini API, Gemini app, AI Mode in Search and Antigravity, with introductory API pricing of $2/$10 per million input/output tokens. Independent evaluations place it level with GPT-6 Astra on Artificial Analysis's Intelligence Index at roughly 60% of Astra's cost per task.

Google's Gemini 4 Argon is now generally available across the paid Gemini API, Google AI Pro and Ultra, AI Mode in Search and Antigravity, at introductory API pricing of $2/$10 per million input/output tokens, with rates reported to double to $4/$20 after November 30.

Dyna Robotics released Dyna-2.1, a physical robot agent built on its Dyna-2 world-action model, which the company says achieved a 99.4% success rate folding restaurant napkins for 24 hours straight. Co-founder Jason Ma detailed the system's architecture and deployments in an interview.