Tracking AI & Technology With Caduceus: 🔗 https://xitter.cf/t.co/xPOhLouCXA
@caduceusvisuals
Joined September 2021
- Tweets11.9K
- Following292
- Followers409
- Likes13.6K
Pinned Tweet
The Fascinating Story of DeepSeek
DeepSeek’s story began with DeepSeek-Coder (2 November 2023) for programming and DeepSeek LLM (29 November 2023) for general conversation.
It soon turned toward harder problems. DeepSeekMath (introduced 5 February 2024) brought GRPO: a training method that compares groups of answers to help a model improve its mathematical reasoning.
This became an important foundation for its later reasoning models. deepseek-ai/deepseek-coder-33b-instruct
During 2024, DeepSeek focused on getting more capability from less computation. V2 (6 May 2024) combined specialist “expert” networks with Multi-head Latent Attention, which sharply reduced the memory needed to generate answers.
Coder-V2 (17 June 2024) expanded programming support to 338 languages. V2.5 (5 September 2024) brought coding and general conversation together. Then V3 (26 December 2024) scaled to 671 billion parameters while activating only 37 billion per token, using FP8 training and multi-token prediction to improve efficiency. GitHub
The next major step was reasoning. R1 and R1-Zero (20 January 2025) explored how reinforcement learning could teach models to work through difficult problems. R1-Zero developed reasoning behaviours without an initial supervised fine-tuning stage.
R1 added training that improved readability and usability, while six smaller distilled models made its reasoning approach more accessible. V3.1 (21 August 2025) later combined thinking and ordinary responses in one model.
V3.2-Exp (29 September 2025) introduced Sparse Attention for more efficient long-context processing, followed by V3.2 and V3.2-Speciale (1 December 2025), with V3.2 integrating reasoning directly into tool use. deepseek-ai/DeepSeek-R1 · GitHub
Alongside its main models, DeepSeek built a wider ecosystem. VL2 (13 December 2024) understood images, charts and documents.
Janus-Pro (27 January 2025) combined image understanding with image generation.
Prover-V2 (30 April 2025) produced mathematical proofs checked by Lean 4, while Math-V2 (27 November 2025) trained models to identify and repair weaknesses in their own proofs.
DeepSeek-OCR (20 October 2025) and OCR 2 (27 January 2026) explored reading documents through compact visual representations. GitHub
In 2026, the focus shifted toward longer context and more capable agents. V4-Pro and V4-Flash entered public preview on 24 April 2026, both supporting one million tokens of context. V4.1-Flash (10 September 2026) brought native image understanding and a new architecture that used fewer active parameters when reading input than when generating output.
The latest update, V4.1-Flash-1017 (17 October 2026), improved coding, reasoning and visual benchmarks while shortening task completion times, according to DeepSeek.
Its surrounding ecosystem now also includes DeepSeek Harness and its own Search and Fetch APIs for agents.
🤖 Made with AI
Apollo retweeted
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing-ou…
Apollo retweeted
Great application of small language models by Meta.
They find that an 8B model picks tools better than a 402B model when trained for the job.
This work focuses on tool selection for agents with more than ten thousand tools.
At that scale, tool descriptions won't fit in context, and a retriever can only rank candidates without knowing which tools the instruction actually needs. Many systems hand this step to a proprietary LLM.
They train Qwen3 models on a Plan, Retrieve, Select workflow. The model splits the instruction into substeps, the retriever returns candidates for each one, and the model commits one tool per substep. Training is SFT on about 20K curated ToolBench instructions, then RL with a reward that scores how many gold tools are covered and at what rank.
On ToolBench G3, the 8B model goes from 36.2% after SFT to 41.5% after RL, above prompted Llama-4-Maverick.
The trained selector also works with different retrievers and on tool corpora from new domains.
arxiv.org/abs/2610.31483
Chat with Paper: academy.dair.ai/papers/train…
Apollo retweeted
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance work: Accurate modeling with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1-pro
Apollo retweeted
DeepSeek V4.1 Flash 1017 is the new #1 openly-available model in the Artificial Analysis Cyber Index. Other proprietary models may offer better cyber defense performance but are restricted access
@deepseek_ai recently released DeepSeek V4.1 Flash 1017, an update to their flagship open-weights model DeepSeek V4.1 Flash. With an Artificial Analysis Cyber Index score of 57%, DeepSeek V4.1 Flash 1017 is the #1 openly-available model for cyber defense capabilities.
The Artificial Analysis Cyber Index measures capability on agentic cyber defense tasks and is composed of three benchmarks: CWE-Bench-AA, DeepsecBench-AA and CyberGym-E2E-AA. The Cyber Index primarily evaluates models through their publicly accessible APIs, reflecting the cyber capabilities available to developers.
Many models, particularly proprietary models from OpenAI, Anthropic and Google, may have stronger defensive cyber capabilities but encounter safety refusals that reduce their overall scores. GPT-6 Sol (Daybreak Blue, max), available through OpenAI’s trusted-access program, leads the Index, scoring 32 points above its public counterpart driven by fewer refusals. DeepSeek V4.1 Flash 1017 encounters no safety refusals.
Key results:
➤ DeepSeek V4.1 Flash 1017 has the highest score among publicly-available models on the Artificial Analysis Cyber Index and #2 position overall. DeepSeek V4.1 Flash 1017 (max) scores 57% ahead of the previously leading publicly-available models Grok 4.7 (xhigh) and MiMo-V2.6-Pro by 1 point. It places #2 overall, with the trusted-access model GPT-6 Sol (Daybreak Blue, max) leading by 12 points with a score of 69%
➤ DeepSeek V4.1 Flash 1017 improves on its predecessor DeepSeek V4.1 Flash by 16 points. DeepSeek V4.1 Flash 1017 (max)’s 16-point improvement from 41% to 57% is driven by a substantial 53-point improvement on CyberGym-E2E-AA from 23% to 76%. On CWE-Bench-AA and DeepsecBench-AA it scores 69% and 25%, compared to 68% and 31% for DeepSeek V4.1 Flash (max)
➤ With a Cost per Task of $0.42, DeepSeek V4.1 Flash 1017 sits on the Capability vs. Cost Pareto frontier. DeepSeek V4.1 Flash 1017 costs $0.42 per task, more than 2x MiMo-V2.6-Pro (56%) at $0.18 per task but 4x cheaper than GPT-6 Sol (Daybreak Blue, max, 69%) at $1.77 per task
Apollo retweeted
🚀 Introducing Hy-Nomas: An Open Source Free-Form Interactive Video World Model
Key Features:
* Free-form interaction with text prompts and camera trajectories
* Long-horizon generation with world memory
* Lightweight LoRA adaptation for specialized domains such as robotics
* Optimized for Inference on Consumer GPUs
Test it and drop feedback anytime！
🔗 Website: hy-nomas.hy.tencent.com/
💻 GitHub:github.com/Tencent-Hunyuan/H…
🤗 Hugging Face: huggingface.co/tencent/Hy-No…
🎥 Demo: hy-nomas.hy.tencent.com/demo
#Hy#WorldModels #VideoGeneration #GenerativeAI
Apollo retweeted
New on our Frontier Red Team blog: using Claude to formally verify software at large scale.
Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time).
With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks.
Read our research in full here: anthropic.com/research/forma…
Apollo retweeted
Introducing Canva Craft 🚀
Extraordinary image generation is now available to everyone, powered by our first frontier model built specifically for design.
With Craft, Canva is now the world's second most advanced AI lab for image generation on @DesignArena's global leaderboard.
See what it can do 👇
Apollo retweeted
Apollo retweeted
GOOGLE 🔥: A new Agent Compute Engine for Google Cloud has been introduced.
ACE is a new VM type built for AI agents, designed for users to operate at scale.
The system includes Agent Gateway, Agent Registry, Agent Identity, and Agent Observability.
Announced at AGNTCon + MCPCon: the new Agent Compute Engine plus expanded Gemini for business agent governance deliver the control you need to run, govern, and scale your agents.
Learn more about Agent Compute Engine and expanded agent governance → goo.gle/4yFN3yT
Apollo retweeted
JUST IN: SpaceXSI has launched Grok for Startups
You get 6–12 months of credits across Grok, Grok Build, Grok Bot and the API, and every single dollar you spend gets matched 1:1 in free credits next month, for six months straight.
It's time to build. x.ai/startups
Apollo retweeted
An output from the updated Carbon checkpoint 😼
Like I said in my previous post, Carbon is noticeably better than Gemini 4 Argon at 3D (tho still not perfect). It seems like Google is currently focusing on improving that domain before moving on to frontend/UI
Apollo retweeted
🤗 Meet our healthcare-enhanced MoE model, @AntLingAGI Ling-3.0-flash-Sante! Built for real-world healthcare research and workflows, with strengthened capabilities in deep research, multi-agent collaboration, and tool use.
Sante shows a significant improvement on professional healthcare tasks, and provides stronger complex medical reasoning.
📊 Open source and open weights:
🔗 huggingface.co/inclusionAI/L…
🔗 modelscope.cn/models/inclusi…
Apollo retweeted
This summer, we invited METR and Redwood Research to pilot an embedded evaluation of our internal monitoring system, providing deep access to our infrastructure over three weeks.
Their findings helped strengthen safeguards as we expanded the system and will help shape future research collaborations with METR and other third party evaluators.
Apollo retweeted
📢 Introducing WorkForge: a scalable framework that constructs verifiable training environments for long-horizon work agents from real-world resources.
Starting from expert workflows, it builds authentic workspaces, extracts checkable facts, and derives tasks plus dual (programmatic + semantic) verifiers grounded in observable evidence.
🚀 We scale to 16.7K environments across 40 professional domains covering 60 file types, enabling consistent gains on long-horizon work benchmarks as data volume and interaction horizon increase.
📚Paper: arxiv.org/pdf/2610.04906 @zhang_leo54389 @ZoeyyyZhououou @Be1ong1
#Tencent Hy#AIAgents #WorkAgents #AgentTraining
Apollo retweeted
Not all forests are equal, and now we can map the difference. 🌳
Meet Forest Typology (ForTy): a 10m map of natural, planted, and farmed forests to help identify deforestation and protect the forests that matter most. 🧵
Apollo retweeted
Oh my god it's REALLY happening.
SCOOP: Major U.S. banks are detecting unauthorized third-party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, sources told Axios.
The findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves at a velocity faster than financial institutions can respond, multiple industry and bank executives told Axios.
Several banks are now preparing for worst-case and even “day after” scenarios, including how to rapidly secure or restore customer accounts if things go wrong, the people said.
Several bank executives told Axios they're struggling to distinguish between human users and AI agents. One executive said banks also struggle to determine whether agents are authorized by customers or potentially operating without their permission, more akin to rogue agent behavior.
Read my latest for Axios here: axios.com/2026/10/22/ai-agen…
Apollo retweeted
PixText offers two ~0.95B open-weight models for local text refinement directly in pixel space.
🤖 modelscope.ai/models/ATH-Maa…
📄 modelscope.ai/papers/2610.18…
🏆 Strong text correction: Reaches 83.06% glyph pass and 94.66% exact target-character OCR accuracy on TextRefinementBench, ahead of the strongest reported baseline.
⚡ Four-step generation: PixText-MeanFlow delivers fast local edits in four steps, retaining an 81.84% glyph pass rate with under one GPU-second amortized cost per image in the paper's setup.
🖼️ Edit text, not the whole image: Provide regions and replacement text, then generate multiple targets jointly while keeping pixels outside the edited footprints unchanged.
🧩 Two checkpoints: Choose the 4-step MeanFlow model for quick use or the 25-step Teacher model for the higher-step workflow.
📜 Apache 2.0.
Apollo retweeted
Introducing Ling-3.1-flash-1019 🚀
It scores 44 on the Artificial Analysis Intelligence Index, up from 41 for Ling-3.1-flash. Gains in agentic work: AA-Briefcase ~1400→1524 Elo and AutomationBench-AA 62%→68%.
API access and open weights coming soon. Stay tuned.
