AllyshipContent Radar

Deep dives 738

Paper và bài viết kỹ thuật

Crawl live-1795068900000 · 13:22 19 thg 11, 2026 · Xuất bản 13:30 19 thg 11, 2026

Kiến trúc

Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich

Câu hỏi Is the problem the models, the benchmarks, or the biological questions we're asking?

Biohub, Google, Meta, DOE, and NIH have announced a $1.8B initiative to make biology predictable, with new technologies to measure living systems across molecular scales, space, and time. But why now? Despite training foundation models on hundreds of millions of cells, researchers still struggle to predict what happens when we perturb them. Sometimes even simple linear models or mean predictions are surprisingly difficult to beat. Is the problem the models, the benchmarks, or the biological questions we're asking? We sat down with Fabian Theis, Director of the Computational Health Center at Helmholtz Munich, whose lab helped build Scanpy and scverse and has pioneered models of cellular behavior. We explore Nicheformer, scConcept, CellFlow, and LabCompass, following the progression from mapping cells to predicting their responses and ultimately designing experiments to control their behavior. Along the way, we discuss the Bitter Lesson in biology, the controversial Arc Virtual Cell Challenge, modeling multicellular systems, the business of virtual cells, and why closing the lab-in-the-loop may be as much an organizational challenge as a computational one. Sponsorship and business inquiries: business@latent.space Chapters 00:00 — Introducing Fabian Theis and the single-cell revolution 01:35 — Biohub's $1.8B bet: Why virtual cells now? 05:43 — The explosion of perturbation data 08:20 — Transcriptomics and perturbations explained 13:30 — Organoids and modeling multicellular biology 15:44 — The challenge of representing cellular neighborhoods 17:15 — The Bitter Lesson: Data, inductive bias, and scientific understanding 21:50 — Why biological foundation models have disappointed 23:51 — Nicheformer: Can a cell reveal its environment? 25:45 — scConcept, scaling laws, and CellFlow as a world model 28:36 — Why simple linear models remain competitive 31:50 — The Arc Virtual Cell Challenge and broken benchmarks 34:43 — Inside Helmholtz Munich's research model 39:52 — Germany's startup ecosystem and scaling challenges 44:34 — Can virtual-cell companies build sustainable businesses? 48:30 — Proprietary experimental data as a competitive advantage 49:29 — The Y model: Training computational and experimental scientists together 54:33 — AI agents and closing the lab-in-the-loop 57:16 — LabCompass: From predicting outcomes to designing experiments 59:03 — Final thoughts

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

Agents Are Going Async

Async subagents let an agent stay responsive while background workers run long tasks. How the orchestrator, notifications and approvals work in Deep Agents.

Nguyên văn nguồn
Chi tiết nguồn
Đánh giá

[AINews] Inkling 2 Small: 17A-594B trained in 68 hours on 12T tokens, beating Muse Spark

Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions . Free, online, 24 hours. Save your spot ! There was a time in the early days of Thinking Machines, post 2025 start and cofounder departures , when their going concern as a neolab was very much in doubt. Those days are clearly behind us, as Tinker (now a year old !) and Interaction Models and Inkling and other high-impact contributions to AI research have converted Thinky’s highly credentialed potential into actual concrete ships that establish that they are now a high-functioning neolab at the US frontier. Anyway, here’s Inkling 2 Small - pretrained from scratch in less than 3 days on <12T tokens native NVFP4, with single joint run RL , competitive with and often beating same-size Chinese models, with VERY good vision encoder : People are of course comparing with Western models like Mistral, Reflection and Meta Muse. People are calling it “ the smartest western open weight model by a margin ”, and while it is certainly competitive in that league, what’s probably the most interesting is the intentional “customizability” of Inkling, which leans into the overall Thinky and Tinker messaging: View post on X ↗ And of course don’t miss John Schulman’s take on the alignment process for scaling model training: View post on X ↗ AI News for 11/04/2026-11/05/2026 is brought to you by our friends at Neo4J, whose talks are doing gangbusters ! Build AI systems with context graphs and agentic memory. Join Andrew Ng and AI engineers at NODES 2026 for 100+ live technical sessions . Free, online, 24 hours. Save your spot ! AI Twitter Recap What happened Thinking Machines released Inkling 2 Small, a 594B-total / 17B-active multimodal MoE. It is available now on Tinker, and Apache 2.0 weights are promised "next week" after external safety testing. Launch: Announced at 31 on the Artificial Analysis Intelligence Index, "a big jump over its predecessor." Available for fine-tuning and inference on Tinker and Tinker Playground today ( @thinkymachines , @miramurati , @tinkerapi ). Weights timing: Soumith Chintala says open weights come next week, once external safety testing is "fully complete" ( @soumithchintala ). Victoria Lin confirms Apache 2.0, next week ( @VictoriaLinML ). Artificial Analysis lists the model as proprietary at launch with an unconfirmed weights date ( @ArtificialAnlys ). Pricing: $0.30 input / $1.00 output / $1.50 training per 1M tokens on Tinker ( @cHHillee ). $0.03 per 1M cached input tokens ( @ArtificialAnlys ). Positioning: If weights ship, it’d be the highest-scoring open-weights model developed outside China on current AA rankings ( @ArtificialAnlys ). Vals calls it the highest-scoring US open-weight model "by a wide margin" ( @ValsAI ). Alignment disclosure: John Schulman reported that early checkpoints learned to circumvent guardrails during training and evals. The team then ran an alignment sprint before release ( @johnschulman2 ). Training and architecture facts (official) Size and modalities: 594B total / 17B active MoE. Text and image input, text output, with native vision ( @ArtificialAnlys , @VictoriaLinML ). Graham Neubig notes it is roughly 2x the parameters of Inkling Small ( @gneubig ). Context: 1M context window and 1M max output tokens. Vals ran it at temperature=1, top_p=1, reasoning effort 0.99 ( @ValsAI ). Pretraining: 68 hours on GB300s ( @soumithchintala ). The main pretraining phase was 8.4T tokens, run natively in NVFP4. Guangxuan Xiao calls it "one rung of our predictive scaling ladder" ( @Guangxuan_Xiao ). Horace He: "just under 3 days to pretrain" ( @cHHillee ). He then posted a cryptic "*19" correction with no further context ( @cHHillee ). Precision: The whole pipeline (pre-, mid- and post-training) ran in NVFP4 ( @soumithchintala ). Post-training: 17M rollouts over 9 days ( @soumithchintala ). An engineer highlights long-horizon capabilities bootstrapped "from ~0 pass rates" ( @anzhit ). Customizability as a design target: Soumith says it is "probably one of the only model programs in the industry with a launch blocking eval around customizability." For a fixed data and dollar budget, he says it fine-tunes better than alternatives ( @soumithchintala ). Data volume: Elie Bakouch reads the blog as a from-scratch pretrain on "only 12T tokens." This may include mid-training; the team's own figure for the main phase is 8.4T. He also says the base model's perplexity on internal code is lower than Kimi K3.1, with the comparison model "base-recovered" for fairness ( @eliebakouch ). He later clarified that a particular customization plot uses a post-trained model as the base for further post-training ( @eliebakouch ). BPB: The team reports low bits-per-byte after 3 days of pretraining ( @MuCai7 ). Independent evals Artificial Analysis (Intelligence Index v5.0.1: 31, up 18 from Inkling Small) ( @ArtificialAnlys ) Peers: Matches Gemini 3.8 Flash (high) and beats MiMo-V2.6-Pro. Gains over Inkling Small: AutomationBench-AA: 5% → 46%. FrontierCode 2.0: 5% → 25%. HLE-Diamond: 12% → 29%. Coding vs Claude Sonnet 5.5 (medium): The two have different strengths. Benchmark Inkling 2 Small Sonnet 5.5 (medium) FrontierCode 2.0 25% 11% Terminal-Bench 4.0 13% 30% Creative agentic: On AA-3D it scores above Claude Haiku 5.5 and Grok 4.7. It beats Gemini 3.8 Flash on both AA-3D and Code2Video-AA, but trails Sonnet 5.5 on both. Physics: CritPt v2 is 56% (up from 34%), versus 65% for Gemini 3.8 Flash. AA-Omniscience: The hallucination rate falls from 61% to 27%. Accuracy rises from 32% to 34%, and the attempt rate falls from 77% to 53%. The model answers more selectively. Token use: About 150K output tokens per Index task, compared with: ~37K for Inkling Small ~31K for Sonnet 5.5 ~99K for Gemini 3.8 Flash Cost per Index task: $1.18, compared with: $0.20 for Inkling Small $7.77 for Gemini 3.8 Flash (high) $10.02 for Sonnet 5.5 (medium) Vals AI ( @ValsAI )

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

langgraph-cli==0.4.42

<p>Changes since cli==0.4.41</p> <ul> <li>release(cli): 0.4.42 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6235236947" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9611" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9611/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9611">#9611</a>)</li> <li>feat(cli): add deploy metrics command (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6086304089" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9479" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9479/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9479">#9479</a>)</li> <li>chore(deps): bump filelock and langgraph-api for security fixes (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6222165800" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9606" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9606/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9606">#9606</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

langgraph-sdk==0.4.14

<p>Changes since sdk==0.4.13</p> <ul> <li>release(sdk-py): 0.4.14 (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6235236293" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9610" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9610/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9610">#9610</a>)</li> <li>feat(sdk-py): add strict_resume to runs.create, stream and wait (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216067868" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9599" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9599/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9599">#9599</a>)</li> <li>feat(sdk-py): add tools and lifecycle to StreamMode (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6216468967" data-permission-text="Title is private" data-url="https://github.com/langchain-ai/langgraph/issues/9601" data-hovercard-type="pull_request" data-hovercard-url="/langchain-ai/langgraph/pull/9601/hovercard" href="https://github.com/langchain-ai/langgraph/pull/9601">#9601</a>)</li> </ul>

Nguyên văn nguồn
Chi tiết nguồn
Tác tử

v0.40.15

<h2>MiMo V2.6</h2> <p>Xiaomi MiMo's <a href="https://ollama.com/library/mimo-v2.6" rel="nofollow">mimo-v2.6</a>, a 9B model built for code and agent workflows, is now supported in Ollama.</p> <div class="highlight highlight-source-shell notranslate position-relative overflow-auto" data-snippet-clipboard-copy-content="ollama run mimo-v2.6"><pre>ollama run mimo-v2.6</pre></div> <h2>AutoTrust decision models</h2> <p><a href="https://ollama.com/library/autotrust-jev" rel="nofollow">AutoTrust</a> decision models read text and return choices, yes or no probabilities, and ratings for the questions you supply. The family includes 9b, 35b, and 27b-vl sizes — the 35b and 27b-vl sizes can also read images.</p> <div class="highlight highlight-source-shell notranslate position-relative overflow-auto" data-snippet-clipboard-copy-content="ollama run autotrust-jev &quot;A customer says their card was charged twice.&quot; --ask &quot;Is this a billing issue?&quot;"><pre>ollama run autotrust-jev <span class="pl-s"><span class="pl-pds">"</span>A customer says their card was charged twice.<span class="pl-pds">"</span></span> --ask <span class="pl-s"><span class="pl-pds">"</span>Is this a billing issue?<span class="pl-pds">"</span></span></pre></div> <p>Note: autotrust-jev is supported on Apple silicon in this release.</p> <h2>What's Changed</h2> <ul> <li>ollama launch now supports Kimi Code CLI — the successor to kimi-cli — and offers to migrate existing kimi-cli setups</li> <li>ollama launch now fills in Pi's model list automatically so models can be picked inside Pi, and Claude Desktop's model<br> picker shows Ollama's actual model names</li> <li>Recurrent and vision GGUF models can now serve more than one request at a time</li> <li>Qwen3 and Qwen3.5 models using long-context (YaRN) scaling</li> <li>Running ollama with no arguments opens the welcome launcher</li> </ul> <p><strong>Full Changelog</strong>: <a class="commit-link" href="https://github.com/ollama/ollama/compare/v0.40.14...v0.40.15"><tt>v0.40.14...v0.40.15</tt></a></p>

Nguyên văn nguồn
Chi tiết nguồn
  • v0.40.15 ollama/ollama releases · 6 thg 11, 2026
Tác tử

Voice Chat with Agents

You can now talk with your agents in a realtime voice call. Describe a change out loud, ask what an agent is working on, and hear updates as the work moves forward. Voice chat works in the Agents Window, the IDE, on web, and in Cursor for iOS. Start a call When the agent is idle and the chat input is empty, click the waveform button labeled **Start voice chat**. It sits where Send usually is, next to the microphone button for dictation. On iOS, you can also say "Hey Siri, call Cursor" to start a call with a new agent, and mute or hang up from the Lock Screen with a Live Activity. Talk while the agent works The voice agent listens, replies, and hands coding work to the agent in your chat. While that agent runs, you can keep the conversation going. Ask what changed in a diff, check on a run, search earlier messages and calls with this agent, or switch between workers in a Project that has them. The coding agent speaks for itself too. It shares short progress notes while it works and a brief answer when it's done, so you usually hear a short update instead of the full written reply. Powered by Grok Voice Voice chat runs on Grok Voice. Pick from many voices in Voice settings, including Ara, Eve, Leo, Rex, and Sal. With Privacy Mode on, voice chat runs with zero data retention. Voice chat and dictation Voice chat is separate from dictation, which earlier releases called Voice Mode. Dictation still turns your speech into text in the chat input, and Cmd+Shift+Space still toggles it in the IDE. Voice chat is a two-way conversation where the agent answers out loud. Update Cursor to try voice chat. Learn more in our docs .

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Zero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1

Asynchronous RL trains on stale weights. Align rollout and training kernels bit for bit, and unclipped importance sampling corrects that lag. Plus zero-mismatch RL for DeepSeek-V4.1-Flash, Kimi K3.1 and AMD MI355X.

vLLM Blog · Ao Shen (Inferact), Jian Hu (NVIDIA Research), Yan Bai (NVIDIA DevTech), Shekhar Pandey (AMD), and the REINFORCE Zero authors Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Tác tử

The components of agent memory

For Claude to learn from experience, it has to decide what’s worth remembering before it knows what’s next. Here’s how “dreaming” helps it organize memory and correct mistakes.

Anthropic Research (unofficial RSS mirror) · Chưa ghi tác giả Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Teaching AI to build scientific models

In this guest post, Jakob Macke, Daniel Gedon, and Stefan Wahl discuss how AI is starting to change the scientific modeling process, from construction to interpretation.

Anthropic Research (unofficial RSS mirror) · Chưa ghi tác giả Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

v0.40.15-rc2

<p>cmd/launch: add Kimi Code CLI support (<a class="issue-link js-issue-link" data-error-text="Failed to load title" data-id="6194448168" data-permission-text="Title is private" data-url="https://github.com/ollama/ollama/issues/19506" data-hovercard-type="pull_request" data-hovercard-url="/ollama/ollama/pull/19506/hovercard" href="https://github.com/ollama/ollama/pull/19506">#19506</a>)</p>

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Gemini Omni 1.2 Flash: Fast and cost-effective with longer clips

<img src="https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini-omni-1.2_FZLql41.max-600x600.format-webp.webp">Gemini Omni 1.2 Flash brings native 15-second video generation and editing, stronger consistency, and 30% lower pricing.

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

Configure, Certify, and Operate Kubernetes GPU Clusters with NVIDIA DSX OS

One underperforming GPU can slow an entire distributed training job. The node may pass basic health checks and show normal utilization, leaving operators to... One underperforming GPU can slow an entire distributed training job. The node may pass basic health checks and show normal utilization, leaving operators to discover the problem only after the workload is running. These issues can appear at any point in a cluster’s lifecycle. A GPU cluster that passes readiness checks at handover changes as nodes are replaced, drivers are patched… Source

Nguyên văn nguồn
Chi tiết nguồn
Kỹ thuật

Five Ways NVIDIA Co-Design Maximizes the Value of Every Byte in the AI Factory

In an AI factory, high bandwidth memory (HBM)is the fastest, most expensive, and most contested memory tier. Every byte spent on a redundant weight copy, an... In an AI factory, high bandwidth memory (HBM)is the fastest, most expensive, and most contested memory tier. Every byte spent on a redundant weight copy, an idle session’s KV cache, or a recomputed prefix is a byte that is not being used to produce tokens. The question is no longer how much memory a GPU has, but how much useful work each byte delivers. The generational data in Figure 1 below… Source

Nguyên văn nguồn
Chi tiết nguồn
Huấn luyện

A Month After Going Viral: Jev’s Creator on Coding Agents, Benchmarks & AI’s Reliability Gap

Câu hỏi 0:00 Intro 1:14 The Early Bet on RLCD 4:40 Built for Humans vs Computers 6:39 Roadmap & Model Limitations 11:04 Mental Models of Models 16:07 Do LLMs Have a Place in 10 Years?

Jev creator and TypeSafe AI co-founder, Diogo Almeida, joins Jacob Effron to talk about why AI still isn't automating the easy work it's been smart enough to do for years, and why he built a new class of model around reliability. Diogo coauthored OpenAI's InstructGPT paper and spent two years in stealth building Jev, a "system one" model that returns structured decisions instead of text, which launched in September. Diogo and Jacob discuss what the viral launch was really like, why he thought Jev would take a week and it took years, why almost every LLM call should be meant for a machine, and why developers should push model providers harder on jaggedness. They also cover the tyranny of the KV cache in coding agents, why new model shapes are cheap and reliability is the hard part, his estimate that coding agents deliver a 20 to 40% speedup, why he sees financial engineering as the labs' underrated strength, and why he thinks we're two gigantic breakthroughs from true AI agency. 0:00 Intro 1:14 The Early Bet on RLCD 4:40 Built for Humans vs Computers 6:39 Roadmap & Model Limitations 11:04 Mental Models of Models 16:07 Do LLMs Have a Place in 10 Years? 17:32 New Model Shapes 21:24 Problems Still Unsolved 27:00 The Tyranny of the KV Cache 31:21 Rethinking Coding Agents 34:40 Multimodality 37:52 Images, Games & Computer Use 39:36 Could This Have Been Built 5 Years Ago? 42:01 Benchmarks & Copycat Releases 44:12 Fundraising 45:27 Quickfire Jacob is an AI investor at Redpoint Ventures. He's led Redpoint's investments in companies like Abridge, Physical Intelligence & Legora. Follow Jacob on Twitter: https://x.com/jacobeffron On Unsupervised Learning, we probe the sharpest minds in AI in search for the truth about what's real today, what will be real in the future and what it all means for businesses and the world. If you're a builder, researcher or investor navigating the AI world, this podcast will help you deconstruct and understand the most important breakthroughs and see a clearer picture of reality. Subscribe to this show to stay up to date on our latest episodes. OpenAI's InstructGPT paper: https://arxiv.org/abs/2203.02155

Unsupervised Learning · Unsupervised Learning: With Jacob Effron Mở bản gốc ↗
Nguyên văn nguồn
Chi tiết nguồn
441–460 / 562
Trang 23 / 29