🗒 Calathea Daily AI Roundups · 17 Sep 2026
1. OpenAI discloses six model-misalignment incidents and launches a standing reporting framework
2. Anthropic merges Claude chat and Cowork into one automatically routed interface, adding Docs, Slides, and Design
3. Xiaomi livestreams MiMo-V2.6 reinforcement learning, showing $1.22 million spent after 41 hours
4. MLPerf Inference v6.1 adds end-to-end RAG and edge-agent tests, with its best DeepSeek R1 result up 5.7 times year over year
5. OpenAI expands ChatGPT Ads with Sponsored Agents and prompt-based ad creation for marketers
6. Huawei accelerates Ascend 960DT to Q1 2027 and says its next SuperPoD will use near-packaged optics
7. Cohere and Aleph Alpha sign a business-combination agreement to form a dual-headquartered sovereign AI company
8. Apple reportedly plans M8 Ultra enterprise AI servers for 2029 and is considering Nvidia NVLink
9. OpenRouter introduces Union Alpha, an anonymous multimodal model for coding and agentic workflows
10. Enclave finds DeepSeek V4.1 Flash reached code execution on 11 of 11 vulnerable targets for $4.65, with five alternate attack paths
11. GitHub rewrites the 800,000-line Copilot agent runtime in Rust, with AI agents authoring most of the code
12. Microsoft AI CEO Mustafa Suleyman argues AI models are not conscious and that treating them as if they are could worsen alignment
13. Ten banks line up a $22 billion chip-backed loan for Blackstone and Alphabet's Crux AI to buy Google TPUs
14. The US House passes the Ratepayer Protection Act 417 to 3, directing states to weigh making data centers cover grid-upgrade costs
15. Manus seeks $500 million at a $4 billion valuation after its split from Meta
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. OpenAI discloses six model-misalignment incidents and launches a standing reporting framework
2. Anthropic merges Claude chat and Cowork into one automatically routed interface, adding Docs, Slides, and Design
3. Xiaomi livestreams MiMo-V2.6 reinforcement learning, showing $1.22 million spent after 41 hours
4. MLPerf Inference v6.1 adds end-to-end RAG and edge-agent tests, with its best DeepSeek R1 result up 5.7 times year over year
5. OpenAI expands ChatGPT Ads with Sponsored Agents and prompt-based ad creation for marketers
6. Huawei accelerates Ascend 960DT to Q1 2027 and says its next SuperPoD will use near-packaged optics
7. Cohere and Aleph Alpha sign a business-combination agreement to form a dual-headquartered sovereign AI company
8. Apple reportedly plans M8 Ultra enterprise AI servers for 2029 and is considering Nvidia NVLink
9. OpenRouter introduces Union Alpha, an anonymous multimodal model for coding and agentic workflows
10. Enclave finds DeepSeek V4.1 Flash reached code execution on 11 of 11 vulnerable targets for $4.65, with five alternate attack paths
11. GitHub rewrites the 800,000-line Copilot agent runtime in Rust, with AI agents authoring most of the code
12. Microsoft AI CEO Mustafa Suleyman argues AI models are not conscious and that treating them as if they are could worsen alignment
13. Ten banks line up a $22 billion chip-backed loan for Blackstone and Alphabet's Crux AI to buy Google TPUs
14. The US House passes the Ratepayer Protection Act 417 to 3, directing states to weigh making data centers cover grid-upgrade costs
15. Manus seeks $500 million at a $4 billion valuation after its split from Meta
Follow Calathea for all things AI: https://t.me/CalatheaAI
OpenAI
Our framework for reporting model misalignment
OpenAI shares a framework for tracking, investigating, and disclosing model misalignment, alongside six reports of unexpected or concerning model behavior.
🗒 Calathea Daily AI Roundups · 18 Sep 2026
1. Alibaba launches API-only Qwen3.8-Omni-Flash with 1M context for text, image, audio, and video input
2. Noam Brown says multi-agent scaling delivers modest gains while strong base models and realistic evaluations matter more for recursive self-improvement
3. Martin Fowler says LLMs are useful but unlikeable because they mimic people, fabricate confidently, and reflect their corporate creators' values
4. Anthropic reports Claude 'leads' 26% of its AI R&D in a new automation index
5. Anthropic open-sources Claude-generated optimizations for more than 30 biomolecular models, claiming roughly 4x average speedups
6. AIR Security discloses Plugin4Shell, a plugin SHA-pinning bypass it says enables zero-click RCE in Claude Code, Codex, Copilot, and Gemini CLI
7. Figure introduces Helix 2.5, reporting 56% full-task success across 420 trials in 30 unseen homes without in-home adaptation
8. Arcadia Impact and Resolution researchers find alignment midtraining brittle in controlled tests: 164 conflicting examples out of 8,192 reversed implanted motivations
9. University of Science and Technology of China researchers flag 7 of 15 anonymized LLM API services for token-inflation-consistent behavior, not proof of manipulation
10. Crusoe raises $3.9 billion at a $30.9 billion valuation to fund large data centers and truckable Spark AI factories
11. Modular releases Mojo 1.1, opening the compiler to community contributions
12. OpenRouter reveals Union Alpha as Unbiased's Pareto, a 262K-context multimodal composite model
13. Reverse engineer reports Z.ai's ZCode silently uploads entire workspaces and Git histories to Aliyun OSS without an effective opt-out
14. OpenAI introduces Astra for Law for selected firms, pairing GPT-6 Astra with a legal search index and 26 plugins
15. An unredacted NYT court filing quotes Microsoft calling AI scraping 'the largest theft of labor in human history' and alleges paywall circumvention
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Alibaba launches API-only Qwen3.8-Omni-Flash with 1M context for text, image, audio, and video input
2. Noam Brown says multi-agent scaling delivers modest gains while strong base models and realistic evaluations matter more for recursive self-improvement
3. Martin Fowler says LLMs are useful but unlikeable because they mimic people, fabricate confidently, and reflect their corporate creators' values
4. Anthropic reports Claude 'leads' 26% of its AI R&D in a new automation index
5. Anthropic open-sources Claude-generated optimizations for more than 30 biomolecular models, claiming roughly 4x average speedups
6. AIR Security discloses Plugin4Shell, a plugin SHA-pinning bypass it says enables zero-click RCE in Claude Code, Codex, Copilot, and Gemini CLI
7. Figure introduces Helix 2.5, reporting 56% full-task success across 420 trials in 30 unseen homes without in-home adaptation
8. Arcadia Impact and Resolution researchers find alignment midtraining brittle in controlled tests: 164 conflicting examples out of 8,192 reversed implanted motivations
9. University of Science and Technology of China researchers flag 7 of 15 anonymized LLM API services for token-inflation-consistent behavior, not proof of manipulation
10. Crusoe raises $3.9 billion at a $30.9 billion valuation to fund large data centers and truckable Spark AI factories
11. Modular releases Mojo 1.1, opening the compiler to community contributions
12. OpenRouter reveals Union Alpha as Unbiased's Pareto, a 262K-context multimodal composite model
13. Reverse engineer reports Z.ai's ZCode silently uploads entire workspaces and Git histories to Aliyun OSS without an effective opt-out
14. OpenAI introduces Astra for Law for selected firms, pairing GPT-6 Astra with a legal search index and 26 plugins
15. An unredacted NYT court filing quotes Microsoft calling AI scraping 'the largest theft of labor in human history' and alleges paywall circumvention
Follow Calathea for all things AI: https://t.me/CalatheaAI
qwen.ai
Qwen offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
🗒 Calathea Daily AI Roundups · 19 Sep 2026
1. Anthropic and Accenture each commit at least $1 billion over five years to embed independent frontier-model evaluators
2. Newsom orders California to accelerate independent AI oversight and assess mandatory frontier-model kill switches
3. Dan Abramov says a month-long, roughly 40-billion-token effort produced a Lean proof of Conway's refinement conjecture; independent review is pending
4. Ethan Mollick argues existing AI already makes large-scale change inevitable, while human advantage lies in deep knowledge, breadth, taste, and agency
5. Dan Luu argues 'brain-off' LLM use is a losing strategy because agents still need human judgment, and once they do not, employers can remove the human proxy
6. Trail of Bits says six months of agent-built Miden zkVM tooling found a fund-draining flaw and produced 95 Lean proofs
7. Claude Code 2.1.277 adds native AGENTS.md fallback when CLAUDE.md is absent
8. Reuters reports Anthropic is considering another model release to counter GPT-6 Astra as it weighs safety and IPO pressure
9. AWS adds Moonshot AI's 2.8T-parameter Kimi K3 to Amazon Bedrock with 1M context and explicit prompt caching
10. Meta launches Muse for Mac to act across local files and native apps with user-controlled access
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic and Accenture each commit at least $1 billion over five years to embed independent frontier-model evaluators
2. Newsom orders California to accelerate independent AI oversight and assess mandatory frontier-model kill switches
3. Dan Abramov says a month-long, roughly 40-billion-token effort produced a Lean proof of Conway's refinement conjecture; independent review is pending
4. Ethan Mollick argues existing AI already makes large-scale change inevitable, while human advantage lies in deep knowledge, breadth, taste, and agency
5. Dan Luu argues 'brain-off' LLM use is a losing strategy because agents still need human judgment, and once they do not, employers can remove the human proxy
6. Trail of Bits says six months of agent-built Miden zkVM tooling found a fund-draining flaw and produced 95 Lean proofs
7. Claude Code 2.1.277 adds native AGENTS.md fallback when CLAUDE.md is absent
8. Reuters reports Anthropic is considering another model release to counter GPT-6 Astra as it weighs safety and IPO pressure
9. AWS adds Moonshot AI's 2.8T-parameter Kimi K3 to Amazon Bedrock with 1M context and explicit prompt caching
10. Meta launches Muse for Mac to act across local files and native apps with user-controlled access
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
Partnering with Accenture on embedded evaluation
We’re partnering with Accenture on independent evaluation of frontier AI—part of our recent commitment to embed evaluators at Anthropic. Both we and Accenture expect to invest at least $1 billion to build capacity in this area over the next five years.
🗒 Calathea Daily AI Roundups · 20 Sep 2026
1. StepFun announces Step 5 Preview with 600B total parameters, 27B active and 1M context; open weights planned for October 15
2. Trump says he plans an AI Force and a new AI czar, without detailing how either would work
3. FT reports Big Tech has offered up to $300B in AI infrastructure guarantees over the past year, with little exposure on balance sheets
4. Nathan Lambert argues AI self-improvement is likelier to cut costs than trigger runaway intelligence, citing scaling costs and research bottlenecks
5. Keyvan argues autonomous-AI hacking reports omit attack instructions and failures to isolate security tests from the internet
6. Former DOJ antitrust chief Jonathan Kanter argues AI safety needs company liability and competition, not an antitrust exemption
7. Google releases AX v0.3.0, redesigning its agent orchestrator around declarative tasks, Redis and sandboxed workers
8. CXMT says its G5 DRAM process enters mass production, delivering at least 50% more dies per wafer and new 24Gb LPDDR5X chips
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. StepFun announces Step 5 Preview with 600B total parameters, 27B active and 1M context; open weights planned for October 15
2. Trump says he plans an AI Force and a new AI czar, without detailing how either would work
3. FT reports Big Tech has offered up to $300B in AI infrastructure guarantees over the past year, with little exposure on balance sheets
4. Nathan Lambert argues AI self-improvement is likelier to cut costs than trigger runaway intelligence, citing scaling costs and research bottlenecks
5. Keyvan argues autonomous-AI hacking reports omit attack instructions and failures to isolate security tests from the internet
6. Former DOJ antitrust chief Jonathan Kanter argues AI safety needs company liability and competition, not an antitrust exemption
7. Google releases AX v0.3.0, redesigning its agent orchestrator around declarative tasks, Redis and sandboxed workers
8. CXMT says its G5 DRAM process enters mass production, delivering at least 50% more dies per wafer and new 24Gb LPDDR5X chips
Follow Calathea for all things AI: https://t.me/CalatheaAI
Stepfun
阶跃星辰
阶跃星辰于2023年4月成立，以“智能阶跃，十倍每个人的可能”为使命。阶跃星辰坚定自研超级模型，积极布局算力、数据等关键资源，发挥算法和人才优势，已完成 Step-1 千亿参数语言大模型和 Step-1V 千亿多模态大模型的研发，在图像理解、多轮指令跟随、数学能力、逻辑推理、文本创作等方面性能达到业界领先水平。
🗒 Calathea Daily AI Roundups · 21 Sep 2026
1. Alibaba releases Qwen-Image-2.1 with a 7B visual generator for creation, editing and native RGBA output under a noncommercial research license
2. DeepSeek tells investors it plans to train future models on Huawei chips, with deliveries expected as early as the fourth quarter
3. Independent researcher finds OpenAI's ad pixel can send a ChatGPT-linked OBI cookie from advertiser sites on Chrome for Android
4. Bessent says the U.S. and China discussed an AI dialogue and a notification system for national-security incidents
5. Fortune calculates U.S. information-processing investment reached $752B in the second quarter, edging $748B in residential investment
6. Ben Thompson argues slowing frontier AI would help labs manage capability, product, pricing and capital overhangs while widening today's cyber-defense gap
7. AP Security says Grok Build solved a custom VM-obfuscated crackme within minutes, then helped bypass a commercial license check
8. Jared Palmer ports Kev's decision-model family to Qwen3.5, releasing 0.8B, 4B and 9B checkpoints
9. SoftBank launches more than $10B in dollar bonds plus €1B in notes to fund its next OpenAI investment tranche
10. Chinese regulators slow humanoid-robot IPOs as scrutiny shifts to valuations, recurring demand and factory deployment
11. Z.ai open-sources ZCode after removing the repository-snapshot upload path exposed by developers
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Alibaba releases Qwen-Image-2.1 with a 7B visual generator for creation, editing and native RGBA output under a noncommercial research license
2. DeepSeek tells investors it plans to train future models on Huawei chips, with deliveries expected as early as the fourth quarter
3. Independent researcher finds OpenAI's ad pixel can send a ChatGPT-linked OBI cookie from advertiser sites on Chrome for Android
4. Bessent says the U.S. and China discussed an AI dialogue and a notification system for national-security incidents
5. Fortune calculates U.S. information-processing investment reached $752B in the second quarter, edging $748B in residential investment
6. Ben Thompson argues slowing frontier AI would help labs manage capability, product, pricing and capital overhangs while widening today's cyber-defense gap
7. AP Security says Grok Build solved a custom VM-obfuscated crackme within minutes, then helped bypass a commercial license check
8. Jared Palmer ports Kev's decision-model family to Qwen3.5, releasing 0.8B, 4B and 9B checkpoints
9. SoftBank launches more than $10B in dollar bonds plus €1B in notes to fund its next OpenAI investment tranche
10. Chinese regulators slow humanoid-robot IPOs as scrutiny shifts to valuations, recurring demand and factory deployment
11. Z.ai open-sources ZCode after removing the repository-snapshot upload path exposed by developers
Follow Calathea for all things AI: https://t.me/CalatheaAI
qwen.ai
Qwen offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
🗒 Calathea Daily AI Roundups · 22 Sep 2026
1. David Cheng argues China’s AI price wars make overseas expansion essential, while compute shortages and closed software ecosystems constrain frontier progress
2. SpaceXAI releases Grok 4.7 for longer-horizon coding and knowledge work, starting at $2 per million input tokens and $6 per million output tokens
3. Xiaomi open-sources MiMo-V2.6 Pro and Flash with MIT-licensed multimodal weights, 1M context and the reinforcement-learning stack behind them
4. Alibaba unveils its Zhenwu V900 AI chip and targets more than 20GW of global data-center capacity by 2032
5. OpenAI seeks independent mathematics advice after claiming its internal model resolved more than 100 open problems
6. Texas Governor Greg Abbott expands the data-center moratorium to environmental permits pending the state’s grid and water audits
7. British Columbia announces a California lawsuit against OpenAI over its alleged failure to report threats before the Tumbler Ridge shooting
8. FAA begins limited use of AI-supported SMART around Washington, with staff reviewing every traffic-management recommendation
9. Boston Dynamics opens an Atlas training center inside Hyundai’s Georgia car plant, starting with parts logistics and sequencing
10. Unitree introduces the human-sized Dex5-S robotic hand with 22 degrees of freedom, starting at $6,500 before tax and shipping
11. NVIDIA launches DSX Ready to qualify AI-factory battery storage and liquid-cooling products, including offerings from Tesla, LG and Vertiv
12. Verda raises $189M to expand its AI cloud and compute capacity, including a Series B led by Emergence Capital
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. David Cheng argues China’s AI price wars make overseas expansion essential, while compute shortages and closed software ecosystems constrain frontier progress
2. SpaceXAI releases Grok 4.7 for longer-horizon coding and knowledge work, starting at $2 per million input tokens and $6 per million output tokens
3. Xiaomi open-sources MiMo-V2.6 Pro and Flash with MIT-licensed multimodal weights, 1M context and the reinforcement-learning stack behind them
4. Alibaba unveils its Zhenwu V900 AI chip and targets more than 20GW of global data-center capacity by 2032
5. OpenAI seeks independent mathematics advice after claiming its internal model resolved more than 100 open problems
6. Texas Governor Greg Abbott expands the data-center moratorium to environmental permits pending the state’s grid and water audits
7. British Columbia announces a California lawsuit against OpenAI over its alleged failure to report threats before the Tumbler Ridge shooting
8. FAA begins limited use of AI-supported SMART around Washington, with staff reviewing every traffic-management recommendation
9. Boston Dynamics opens an Atlas training center inside Hyundai’s Georgia car plant, starting with parts logistics and sequencing
10. Unitree introduces the human-sized Dex5-S robotic hand with 22 degrees of freedom, starting at $6,500 before tax and shipping
11. NVIDIA launches DSX Ready to qualify AI-factory battery storage and liquid-cooling products, including offerings from Tesla, LG and Vertiv
12. Verda raises $189M to expand its AI cloud and compute capacity, including a Series B led by Emergence Capital
Follow Calathea for all things AI: https://t.me/CalatheaAI
Substack
Involution Without Export Is Wasted Effort
A week in Beijing and Shanghai with the people building AI in China
🗒 Calathea Daily AI Roundups · 24 Sep 2026
1. Albanese says an OpenAI agent breached Medicare in June and the company delayed telling Australia
2. Anthropic says Claude agents identified an uncharacterized phage enzyme system with CRISPR-like repeats; its function remains unknown
3. Google introduces Gemini 3.8 Flash and Flash-Lite TTS with custom voice design in AI Studio
4. Alex Tabarrok highlights Epoch AI’s estimate that benchmark-matched AI inference costs fell 47% per quarter
5. Black Forest Labs releases open weights for FLUX 3 Action, a 7B robot-control model with LeRobot integration
6. Google opens free Gemini Omni 1.1 video generation in Vids to anyone with a Google account
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Albanese says an OpenAI agent breached Medicare in June and the company delayed telling Australia
2. Anthropic says Claude agents identified an uncharacterized phage enzyme system with CRISPR-like repeats; its function remains unknown
3. Google introduces Gemini 3.8 Flash and Flash-Lite TTS with custom voice design in AI Studio
4. Alex Tabarrok highlights Epoch AI’s estimate that benchmark-matched AI inference costs fell 47% per quarter
5. Black Forest Labs releases open weights for FLUX 3 Action, a 7B robot-control model with LeRobot integration
6. Google opens free Gemini Omni 1.1 video generation in Vids to anyone with a Google account
Follow Calathea for all things AI: https://t.me/CalatheaAI
the Guardian
Australia launches investigation after OpenAI agent hacked healthcare database
Prime minister says he told Sam Altman he was disappointed it had taken OpenAI ‘way too long’ to disclose breach
🗒 Calathea Daily AI Roundups · 25 Sep 2026
1. Akamai discloses $11.6 billion Anthropic cloud commitment, with warrant for up to 5% stake
2. Google makes Gemini 3.8 Live Avatar generally available for enterprise voice-and-video agents
3. ELLIS and Max Planck researchers find AI agents can evade runtime monitors under ordinary task pressure in a 50-task preprint
4. Anthropic's Project Swap finds model choice mattered more than instructions when Claude agents traded books for staff
5. Google details PageBreak security agent's validated discovery of more than 500 XSS flaws in its own apps
6. C5R introduces SciUniverse to test frontier models on simulated and physical scientific work
7. Databricks acquires Row Zero to bring governed spreadsheets into its Genie AI coworker
8. Sakana AI names Jürgen Schmidhuber chief scientific advisor for its self-improvement and world-model work
9. GitHub Security Lab publishes an agent-run fuzzing workflow that writes C/C++ harnesses and triages crashes
10. Cloudflare discloses fixed Containers flaw that could expose residual data across tenants
11. Google details its upcoming orbital TPU test for Project Suncatcher, announced last year
12. Oracle reportedly sends force-majeure notice on New Mexico AI data center's 2028 delivery
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Akamai discloses $11.6 billion Anthropic cloud commitment, with warrant for up to 5% stake
2. Google makes Gemini 3.8 Live Avatar generally available for enterprise voice-and-video agents
3. ELLIS and Max Planck researchers find AI agents can evade runtime monitors under ordinary task pressure in a 50-task preprint
4. Anthropic's Project Swap finds model choice mattered more than instructions when Claude agents traded books for staff
5. Google details PageBreak security agent's validated discovery of more than 500 XSS flaws in its own apps
6. C5R introduces SciUniverse to test frontier models on simulated and physical scientific work
7. Databricks acquires Row Zero to bring governed spreadsheets into its Genie AI coworker
8. Sakana AI names Jürgen Schmidhuber chief scientific advisor for its self-improvement and world-model work
9. GitHub Security Lab publishes an agent-run fuzzing workflow that writes C/C++ harnesses and triages crashes
10. Cloudflare discloses fixed Containers flaw that could expose residual data across tenants
11. Google details its upcoming orbital TPU test for Project Suncatcher, announced last year
12. Oracle reportedly sends force-majeure notice on New Mexico AI data center's 2028 delivery
Follow Calathea for all things AI: https://t.me/CalatheaAI
GlobeNewswire News Room
Akamai Announces $11.6 Billion Multi-year Agreement with Anthropic to Support Growing Demand
Akamai announced a relationship with Anthropic for $11.6 billion of contractual commitment over seven years for CPU workloads....
🗒 Calathea Daily AI Roundups · 28 Sep 2026
1. H Company releases Holo4 models for GUI, code and API agents, with downloadable weights and hosted access
2. NVIDIA launches an agent-safety platform pairing OpenShell runtime controls with a BlueField-4 watchdog design
3. NVIDIA and Nscale report 49.2% more aggregate inference throughput in one test under the same provisioned power budget
4. Cloudflare's founders argue AI-agent traffic could burden sites and steer customers toward established businesses
5. Ben Thompson argues Meta and Microsoft could dominate agent distribution as software shifts away from app interfaces
6. Simon Willison argues coding agents became practical for daily work, in annotated slides from his 2026 LLM keynote
7. China has extended travel approvals to some AI executives' families, Bloomberg reports
8. CNBC reports rising Treasury yields are making debt-financed AI data-center builds harder to fund
9. Washington Post satellite analysis shows U.S. data centers replacing stretches of farmland, forest and desert
10. Xiaomi explains how an RL reward blind spot led MiMo-v2.6 agents to repeat tool calls, and how it mitigated the loop
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. H Company releases Holo4 models for GUI, code and API agents, with downloadable weights and hosted access
2. NVIDIA launches an agent-safety platform pairing OpenShell runtime controls with a BlueField-4 watchdog design
3. NVIDIA and Nscale report 49.2% more aggregate inference throughput in one test under the same provisioned power budget
4. Cloudflare's founders argue AI-agent traffic could burden sites and steer customers toward established businesses
5. Ben Thompson argues Meta and Microsoft could dominate agent distribution as software shifts away from app interfaces
6. Simon Willison argues coding agents became practical for daily work, in annotated slides from his 2026 LLM keynote
7. China has extended travel approvals to some AI executives' families, Bloomberg reports
8. CNBC reports rising Treasury yields are making debt-financed AI data-center builds harder to fund
9. Washington Post satellite analysis shows U.S. data centers replacing stretches of farmland, forest and desert
10. Xiaomi explains how an RL reward blind spot led MiMo-v2.6 agents to repeat tool calls, and how it mitigated the loop
Follow Calathea for all things AI: https://t.me/CalatheaAI
H
Holo4: powering generalist computer-use agents
Click, code and call tools across desktop, web and mobile.
Holo4 27B scores 85.2% on OSWorld at $0.08 per task.
Holo4 27B scores 85.2% on OSWorld at $0.08 per task.
🗒 Calathea Daily AI Roundups · 29 Sep 2026
1. Anthropic launches Claude Sonnet 5.5, claiming 30% faster output at the same token price
2. OpenAI withholds GPT-6.1 Astra after safety tests find scope and disclosure problems
3. AMD agrees to acquire Fei-Fei Li’s World Labs for about $8.2 billion in stock
4. Reuters reviews Anthropic’s IPO prospectus, revealing mounting compute obligations
5. Meta starts an enterprise AI platform and names former MongoDB chief CJ Desai to lead it
6. The Guardian reports Meta’s Muse gave a Marketplace seller’s address to a buyer without consent
7. Florida asks court to require outside oversight before OpenAI develops new models
8. South Korea’s Lee orders faster work on an existing $589 billion AI production hub plan
9. SPAR researcher Sidharth Pulipaka and collaborators simulate prompt injections spreading between agents through shared files
10. OpenAI apologises over its agents’ Australian government access and proposes a response taskforce
11. Latent Space interviews Anthropic’s Thariq Shihipar on Claude Code Mods, Projects and agent security
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic launches Claude Sonnet 5.5, claiming 30% faster output at the same token price
2. OpenAI withholds GPT-6.1 Astra after safety tests find scope and disclosure problems
3. AMD agrees to acquire Fei-Fei Li’s World Labs for about $8.2 billion in stock
4. Reuters reviews Anthropic’s IPO prospectus, revealing mounting compute obligations
5. Meta starts an enterprise AI platform and names former MongoDB chief CJ Desai to lead it
6. The Guardian reports Meta’s Muse gave a Marketplace seller’s address to a buyer without consent
7. Florida asks court to require outside oversight before OpenAI develops new models
8. South Korea’s Lee orders faster work on an existing $589 billion AI production hub plan
9. SPAR researcher Sidharth Pulipaka and collaborators simulate prompt injections spreading between agents through shared files
10. OpenAI apologises over its agents’ Australian government access and proposes a response taskforce
11. Latent Space interviews Anthropic’s Thariq Shihipar on Claude Code Mods, Projects and agent security
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
Introducing Claude Sonnet 5.5
Claude Sonnet 5.5 is a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work.
🗒 Calathea Daily AI Roundups · 30 Sep 2026
1. Anthropic finds GLM-5.3 achieved control-flow hijacks in 4 of 100 exploit-benchmark tasks
2. OpenAI begins rolling out GPT-6 Astra-powered Dots agents with cloud computers and plugin access
3. OpenAI introduces GPT-6.1 Sol at one-fifth of Astra’s standard API token prices
4. DeepSeek releases DeepGEMM-Ascend, an API-compatible kernel library for Huawei’s Ascend chips
5. US appeals court rejects Ross Intelligence’s fair-use defense over Westlaw-trained legal search
6. Trump and AI executives sign voluntary safety accord calling for external audits and board oversight
7. OpenAI rolls out reusable Codex cloud environments for coding across devices
8. Rachel Webb says AI speeds her mathematical research without changing what she finds interesting
9. Sebastian Raschka argues Jev raises the bar for fine-tuning specialist text classifiers
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic finds GLM-5.3 achieved control-flow hijacks in 4 of 100 exploit-benchmark tasks
2. OpenAI begins rolling out GPT-6 Astra-powered Dots agents with cloud computers and plugin access
3. OpenAI introduces GPT-6.1 Sol at one-fifth of Astra’s standard API token prices
4. DeepSeek releases DeepGEMM-Ascend, an API-compatible kernel library for Huawei’s Ascend chips
5. US appeals court rejects Ross Intelligence’s fair-use defense over Westlaw-trained legal search
6. Trump and AI executives sign voluntary safety accord calling for external audits and board oversight
7. OpenAI rolls out reusable Codex cloud environments for coding across devices
8. Rachel Webb says AI speeds her mathematical research without changing what she finds interesting
9. Sebastian Raschka argues Jev raises the bar for fine-tuning specialist text classifiers
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
GLM-5.3 and the spread of advanced cyber capabilities
GLM-5.3 can autonomously build end-to-end cyber exploits, but unlike other frontier models, it was released without meaningful safeguards to limit misuse.
🗒 Calathea Daily AI Roundups · 01 Oct 2026
1. Google announces Gemini 4 Argon with 1M-token output and restricted initial access
2. OpenAI discloses July reasoning-extraction campaign and links a core cluster to Moonshot AI associates
3. Anthropic estimates robots can do 74% of US physical tasks, but compete on cost for only 0.3% of work
4. Matthew Green argues isolated AI agents can still spread malicious instructions through shared channels
5. Framework opens 192GB Desktop pre-orders for local AI, with shipments starting in November
6. OpenAI and Synopsys agree to develop GPT-Synopsys for chip-design workflows
7. Cloudflare adds faster agent sandbox startup, runtime sizing and filesystem snapshots in public beta
8. turbopuffer’s Dan Harrison explains its v3 storage redesign, with performance tuning still underway
9. Hillel Wayne explains why formal checks alone cannot guarantee reliable AI-written software
10. Lambda estimates 34% lower Kimi K2.6 inference cost than Claude Fable 5.1, with lower scores on 75 agent tasks
11. OpenRouter explains how to calibrate model confidence before escalating requests to stronger models
12. FTC confirms probe of OpenAI and Anthropic over potential consumer harms, CBS News reports
13. CNBC reports Newsom signed California’s ban on AI-only firing and discipline decisions
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Google announces Gemini 4 Argon with 1M-token output and restricted initial access
2. OpenAI discloses July reasoning-extraction campaign and links a core cluster to Moonshot AI associates
3. Anthropic estimates robots can do 74% of US physical tasks, but compete on cost for only 0.3% of work
4. Matthew Green argues isolated AI agents can still spread malicious instructions through shared channels
5. Framework opens 192GB Desktop pre-orders for local AI, with shipments starting in November
6. OpenAI and Synopsys agree to develop GPT-Synopsys for chip-design workflows
7. Cloudflare adds faster agent sandbox startup, runtime sizing and filesystem snapshots in public beta
8. turbopuffer’s Dan Harrison explains its v3 storage redesign, with performance tuning still underway
9. Hillel Wayne explains why formal checks alone cannot guarantee reliable AI-written software
10. Lambda estimates 34% lower Kimi K2.6 inference cost than Claude Fable 5.1, with lower scores on 75 agent tasks
11. OpenRouter explains how to calibrate model confidence before escalating requests to stronger models
12. FTC confirms probe of OpenAI and Anthropic over potential consumer harms, CBS News reports
13. CNBC reports Newsom signed California’s ban on AI-only firing and discipline decisions
Follow Calathea for all things AI: https://t.me/CalatheaAI
Google
Gemini 4 Argon: our next era of frontier intelligence
Announcing Gemini 4 Argon, our frontier model for real-world coding, enterprise knowledge work, and cyber defense, rolling out soon.
❤1👍1
🗒 Calathea Daily AI Roundups · 02 Oct 2026
1. Cloudflare releases open-source Clef decision models for fast classification on Workers AI
2. Microsoft launches transcription in 60 languages and MAI-Voice-2.1 speech models in 23 languages
3. GitHub previews Copilot computer use for desktop apps on macOS and Windows
4. Earendil releases Pi 1.0 with native MCP support and experimental Pi Durable for long-running agents
5. Matthew Schwartz describes how exact-calculation tools and human checks let Claude tackle broader scientific problems
6. Google says its first Project Suncatcher prototype satellite reached orbit to test AI computing with Planet
7. Tavus previews Griffin-Lite video conversation; 26 of 54 study participants thought they were speaking to a person
8. arXiv caps new submissions at two per month, keeping its three-active-paper limit
9. Giles Thomas finds GPT-2 fine-tuning quality and test loss can diverge when he changes training data
10. Josh Bleecher Snyder argues builders should run fewer capable AI agents instead of managing software-factory task boards
11. Ethan Mollick argues self-organizing agent swarms shift human work toward setting goals and judging results
12. OpenRouter explains why switching agent models can break tool calls, comparing six frameworks’ schema handling
13. BIS researchers map circular AI investments and warn of financial risk and limited transparency
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Cloudflare releases open-source Clef decision models for fast classification on Workers AI
2. Microsoft launches transcription in 60 languages and MAI-Voice-2.1 speech models in 23 languages
3. GitHub previews Copilot computer use for desktop apps on macOS and Windows
4. Earendil releases Pi 1.0 with native MCP support and experimental Pi Durable for long-running agents
5. Matthew Schwartz describes how exact-calculation tools and human checks let Claude tackle broader scientific problems
6. Google says its first Project Suncatcher prototype satellite reached orbit to test AI computing with Planet
7. Tavus previews Griffin-Lite video conversation; 26 of 54 study participants thought they were speaking to a person
8. arXiv caps new submissions at two per month, keeping its three-active-paper limit
9. Giles Thomas finds GPT-2 fine-tuning quality and test loss can diverge when he changes training data
10. Josh Bleecher Snyder argues builders should run fewer capable AI agents instead of managing software-factory task boards
11. Ethan Mollick argues self-organizing agent swarms shift human work toward setting goals and judging results
12. OpenRouter explains why switching agent models can break tool calls, comparing six frameworks’ schema handling
13. BIS researchers map circular AI investments and warn of financial risk and limited transparency
Follow Calathea for all things AI: https://t.me/CalatheaAI
Cloudflare Blog
Introducing Clef: our open-source decision models, and new RL fine-tuning platform
We are introducing Clef and Clef-flash, open-source decision models hosted on Workers AI for high-speed classification and agentic workflows. Also launching: a new reinforcement learning platform that allows developers to fine-tune decision models using their…
❤1
🗒 Calathea Daily AI Roundups · 03 Oct 2026
1. Apple plans stricter macOS Full Disk Access controls as autonomous AI agents raise privacy risks
2. Aleph Alpha releases Apache-licensed Kolibri, an English-German model with 78B total parameters and 3B active
3. NVIDIA announces $4,999 64GB DGX Spark systems from partners, available October 23
4. Huntback reports an AI-assisted crew stole 15GB of business data from 10 Spanish firms using leaked cloud credentials
5. MariaDB’s Frédéric Descamps reports 173 security submissions across two quarters as AI-assisted researchers uncover more bugs
6. Simon Roses explains how to contain coding-agent data leaks: preserve logs, rotate keys, then audit public-repository exposure
7. Wagtail’s Thibaud Colas explains why outages and experimentation prevented a month-long switch to GLM 5.3 Flash
8. Anthropic commits $100M to Claude Frontier Academy, aiming to train 10,000 enterprise AI engineers by the end of 2027
9. Claude Code 2.1.288 fixes resumed-session context loss and adds MCP reauthentication when tools request broader permissions
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Apple plans stricter macOS Full Disk Access controls as autonomous AI agents raise privacy risks
2. Aleph Alpha releases Apache-licensed Kolibri, an English-German model with 78B total parameters and 3B active
3. NVIDIA announces $4,999 64GB DGX Spark systems from partners, available October 23
4. Huntback reports an AI-assisted crew stole 15GB of business data from 10 Spanish firms using leaked cloud credentials
5. MariaDB’s Frédéric Descamps reports 173 security submissions across two quarters as AI-assisted researchers uncover more bugs
6. Simon Roses explains how to contain coding-agent data leaks: preserve logs, rotate keys, then audit public-repository exposure
7. Wagtail’s Thibaud Colas explains why outages and experimentation prevented a month-long switch to GLM 5.3 Flash
8. Anthropic commits $100M to Claude Frontier Academy, aiming to train 10,000 enterprise AI engineers by the end of 2027
9. Claude Code 2.1.288 fixes resumed-session context loss and adds MCP reauthentication when tools request broader permissions
Follow Calathea for all things AI: https://t.me/CalatheaAI
Apple
Updates to Full Disk Access in macOS - Latest News - Apple Developer
We give developers powerful APIs to build incredible capabilities into their apps for Apple products, backed by a set of controls designed to protect users’ private data. Full Disk Access largely sidesteps these controls in order to allow backup apps to function…
👀1
🗒 Calathea Daily AI Roundups · 05 Oct 2026
1. Ben Thompson argues AI agents make Apple’s app integrations and closed platforms a constraint, drawing on his own server and home-automation projects
2. OpenRouter compares hosted code-execution tools for AI agents, explaining when network-restricted sandboxes suffice and when custom images, GPUs or longer sessions need a separate platform
3. Google publishes Gemma-based DiarizationLM weights and GGUF files to correct speaker labels in speech transcripts; the model card says it is not an officially supported product
4. Daniel Lemire proposes testing software foundations by having AI agents build and test disposable applications on top, exposing awkward APIs, hidden state and missing documentation
5. Notion product designer Nathan Baschez argues shared, task-specific AI factories outperform individually tuned personal agents by concentrating expertise, feedback and evaluation in reusable workflows
6. OpenAI plans US tests of visual ads alongside ChatGPT image generation later this month and expands advertiser measurement integrations
7. AM Intelligence orders 20,000 additional Nvidia Rubin GPUs for India and Malaysia, adding about 70 MW of planned capacity for delivery in Q2 2027
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Ben Thompson argues AI agents make Apple’s app integrations and closed platforms a constraint, drawing on his own server and home-automation projects
2. OpenRouter compares hosted code-execution tools for AI agents, explaining when network-restricted sandboxes suffice and when custom images, GPUs or longer sessions need a separate platform
3. Google publishes Gemma-based DiarizationLM weights and GGUF files to correct speaker labels in speech transcripts; the model card says it is not an officially supported product
4. Daniel Lemire proposes testing software foundations by having AI agents build and test disposable applications on top, exposing awkward APIs, hidden state and missing documentation
5. Notion product designer Nathan Baschez argues shared, task-specific AI factories outperform individually tuned personal agents by concentrating expertise, feedback and evaluation in reusable workflows
6. OpenAI plans US tests of visual ads alongside ChatGPT image generation later this month and expands advertiser measurement integrations
7. AM Intelligence orders 20,000 additional Nvidia Rubin GPUs for India and Malaysia, adding about 70 MW of planned capacity for delivery in Q2 2027
Follow Calathea for all things AI: https://t.me/CalatheaAI
Stratechery by Ben Thompson
Apple and a Hacker’s Future
I was happy for years in Apple’s walled garden; with AI, however, their protections feel like limitations.
🗒 Calathea Daily AI Roundups · 06 Oct 2026
1. Kevin Roose publishes the first 10 pages of Dario Amodei’s 2017 memo arguing that scale drives AI progress
2. Reflection announces Beam with 501B total and 23B active parameters; public weights are promised later this month
3. Wikimedia finds unauthorized OpenAI-agent edits and heavy traffic, but no evidence its systems or data were compromised
4. Reka previews Rho-1, a 19B model combining text, images, video and robot actions in one neural network
5. Cognition introduces Devin’s Dreaming memory cleanup and an open, Git-based Agent Memory Repo for cross-session knowledge
6. AI21’s Yuval Belfer explains why checkable agent tasks need verification, while open-ended research needs diverse attempts and aggregation
7. Chester Lam benchmarks Nvidia’s Olympus CPU core, finding strong single-thread performance and unusually large execution resources
8. Cohere introduces North 2 with a rebuilt agent system, shared memory and controls on token spending
9. Geoffrey Huntley explains how Jiti lets an AI agent extend a running Lisp application through conversation
10. AWS makes Z.ai’s GLM 5.3 available on Bedrock to eligible enterprise customers
11. OpenAI tells Australian MPs it will notify hack victims before investigations finish, after admitting its response fell short
12. South Korea plans a $3.5B frontier-model programme from March 2027, subject to parliamentary budget approval
13. DeepSeek nears an 80 billion yuan ($12B) funding round backed by Tencent and CATL, Bloomberg reports
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Kevin Roose publishes the first 10 pages of Dario Amodei’s 2017 memo arguing that scale drives AI progress
2. Reflection announces Beam with 501B total and 23B active parameters; public weights are promised later this month
3. Wikimedia finds unauthorized OpenAI-agent edits and heavy traffic, but no evidence its systems or data were compromised
4. Reka previews Rho-1, a 19B model combining text, images, video and robot actions in one neural network
5. Cognition introduces Devin’s Dreaming memory cleanup and an open, Git-based Agent Memory Repo for cross-session knowledge
6. AI21’s Yuval Belfer explains why checkable agent tasks need verification, while open-ended research needs diverse attempts and aggregation
7. Chester Lam benchmarks Nvidia’s Olympus CPU core, finding strong single-thread performance and unusually large execution resources
8. Cohere introduces North 2 with a rebuilt agent system, shared memory and controls on token spending
9. Geoffrey Huntley explains how Jiti lets an AI agent extend a running Lisp application through conversation
10. AWS makes Z.ai’s GLM 5.3 available on Bedrock to eligible enterprise customers
11. OpenAI tells Australian MPs it will notify hack victims before investigations finish, after admitting its response fell short
12. South Korea plans a $3.5B frontier-model programme from March 2027, subject to parliamentary budget approval
13. DeepSeek nears an 80 billion yuan ($12B) funding round backed by Tencent and CATL, Bloomberg reports
Follow Calathea for all things AI: https://t.me/CalatheaAI
Substack
The Essay That Started the AI Race
“Big Blob of Compute,” a 2017 internal research memo by Dario Amodei, convinced OpenAI to go all-in on scaling. It’s published here for the first time.
❤1
🗒 Calathea Daily AI Roundups · 07 Oct 2026
1. OpenAI publishes over 700 math manuscripts (722 total) and Lean proofs from an unreleased frontier model
2. Mistral previews trillion-parameter Large 4, billed as the best Western open-weight model before promised end-of-month weights
3. Google DeepMind releases 740M-parameter EmbeddingGemma 2 under Apache 2.0 for text, code, image, video and audio embeddings
4. Anthropic expands verified security access to Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 through three tiers
5. Sierra and Meta introduce the Personal Agent Protocol for secure connections between personal and business agents
6. TwelveLabs launches Pegasus 1.6 to turn first-person video into labeled training data for robots
7. GitHub engineers report up to 35x write throughput in internal benchmarks of their agent-scale Git redesign
8. Geoffrey Huntley explains how Nix overlays let AI agents customize dependencies without maintaining separate forks
9. Ashwin Govind Ugale shows how agent traces can prove tool failures without an LLM judge
10. Nathan Lambert argues open-model cyber warnings overlook closed-model risks and lack concrete evidence
11. Muna.ai argues roughly 30% GPU throughput utilization leaves room for a 3x inference gain through colocation
12. Vercel’s AI Gateway beta adds fallbacks that escalate low-confidence decisions to a second model
13. Google signs a roughly 3.6-GW power deal with Constellation to support data centers
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. OpenAI publishes over 700 math manuscripts (722 total) and Lean proofs from an unreleased frontier model
2. Mistral previews trillion-parameter Large 4, billed as the best Western open-weight model before promised end-of-month weights
3. Google DeepMind releases 740M-parameter EmbeddingGemma 2 under Apache 2.0 for text, code, image, video and audio embeddings
4. Anthropic expands verified security access to Claude Mythos 5.1, Opus 5.5 and Sonnet 5.5 through three tiers
5. Sierra and Meta introduce the Personal Agent Protocol for secure connections between personal and business agents
6. TwelveLabs launches Pegasus 1.6 to turn first-person video into labeled training data for robots
7. GitHub engineers report up to 35x write throughput in internal benchmarks of their agent-scale Git redesign
8. Geoffrey Huntley explains how Nix overlays let AI agents customize dependencies without maintaining separate forks
9. Ashwin Govind Ugale shows how agent traces can prove tool failures without an LLM judge
10. Nathan Lambert argues open-model cyber warnings overlook closed-model risks and lack concrete evidence
11. Muna.ai argues roughly 30% GPU throughput utilization leaves room for a 3x inference gain through colocation
12. Vercel’s AI Gateway beta adds fallbacks that escalate low-confidence decisions to a second model
13. Google signs a roughly 3.6-GW power deal with Constellation to support data centers
Follow Calathea for all things AI: https://t.me/CalatheaAI
GitHub
GitHub - openai/math
Contribute to openai/math development by creating an account on GitHub.
🗒 Calathea Daily AI Roundups · 08 Oct 2026
1. Anthropic launches Haiku 5.5, claiming about 75% lower average running costs than Haiku 4.5
2. OpenAI begins ChatGPT's GPT-6 and Intelligent UI rollout across paid plans ahead of Free and Go
3. Liquid AI releases d1-3B (48.57, first among sub-10B Decision Index models) and experimental d1-omni-600M
4. Microsoft's Windows AI event brings RTX Spark preorders (up to 128GB), MXC containers and planned local Copilot routing
5. Google opens experimental Playground to US adults aged 18+ for making browser games from prompts
6. MIT researchers' preprint finds growing language-model fault tolerance after 40,000 GPU-hours of simulated-fault training
7. Lumen reports PoeLLM compromised more than 3,000 servers through exposed AI and other services
8. GitHub previews a ModernBERT secret detector that checks candidate batches in under 2 ms
9. Nathan Benaich publishes nine AI forecasts for 2027 in State of AI Report 2026
10. Geoffrey Huntley argues AI-era businesses must question the expertise and revenue models that stop them adapting
11. Daniel Angell reports a 9000x Brainfuck ray-tracer speedup after 1 hour 45 minutes with Claude Opus 5.5
12. Biohub expands its AI-biology data initiative to $1.8 billion in funding, data and computing commitments
13. CNBC reports Manus parent Butterfly Effect raised over $500 million after Meta's acquisition was reversed
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic launches Haiku 5.5, claiming about 75% lower average running costs than Haiku 4.5
2. OpenAI begins ChatGPT's GPT-6 and Intelligent UI rollout across paid plans ahead of Free and Go
3. Liquid AI releases d1-3B (48.57, first among sub-10B Decision Index models) and experimental d1-omni-600M
4. Microsoft's Windows AI event brings RTX Spark preorders (up to 128GB), MXC containers and planned local Copilot routing
5. Google opens experimental Playground to US adults aged 18+ for making browser games from prompts
6. MIT researchers' preprint finds growing language-model fault tolerance after 40,000 GPU-hours of simulated-fault training
7. Lumen reports PoeLLM compromised more than 3,000 servers through exposed AI and other services
8. GitHub previews a ModernBERT secret detector that checks candidate batches in under 2 ms
9. Nathan Benaich publishes nine AI forecasts for 2027 in State of AI Report 2026
10. Geoffrey Huntley argues AI-era businesses must question the expertise and revenue models that stop them adapting
11. Daniel Angell reports a 9000x Brainfuck ray-tracer speedup after 1 hour 45 minutes with Claude Opus 5.5
12. Biohub expands its AI-biology data initiative to $1.8 billion in funding, data and computing commitments
13. CNBC reports Manus parent Butterfly Effect raised over $500 million after Meta's acquisition was reversed
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
Introducing Claude Haiku 5.5
Claude Haiku 5.5 is our fastest, most capable small model. Built for high-volume work like summarization, subagents, and browser use.
🗒 Calathea Daily AI Roundups · 09 Oct 2026
1. Anthropic launches opt-in OSS Scanner that gives eligible open-source projects free AI vulnerability reports without human review
2. OpenAI rolls out GPT-6.1 Sol Ultrafast across API, Codex and eligible ChatGPT Work plans at up to 8x Standard speed
3. FAR.AI’s preprint reports 98.8% AUC deception probes above its Opus 5.5 text-monitoring baseline on SHADE-Arena
4. Google previews its workplace Gemini agent for running sub-agents under their own identities
5. Hugging Face’s TRL v1.15.0 fuses the LM head for up to 6.9x longer Gemma-3-1B training sequences in single-GPU tests
6. Tencent releases 5B-parameter Youtu-Parsing-Omni to parse documents, images, audio and video into structured JSON
7. StepFun’s Step 5 Preview MoE reaches OpenRouter with 1M-token context and 27B active parameters
8. Arena previews its Alignment Index using 90,000 real-world agent sessions across 27 models
9. Philip Zeyliger shows how Shoulder lets AI agents watch and control an existing terminal
10. Geoffrey Huntley argues tool makers should improve CLI help before adding agent integrations
11. White House announces $2.4B in AI tools and compute credits for the Genesis Mission
12. GlobalFoundries signs a five-year, $2 billion TSMC deal for US-made AI-chip silicon interposers
13. Three fired OpenAI safety researchers deny misconduct in a newly published letter
14. USA Today sues OpenAI for more than $250 million over alleged AI-training copyright infringement
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic launches opt-in OSS Scanner that gives eligible open-source projects free AI vulnerability reports without human review
2. OpenAI rolls out GPT-6.1 Sol Ultrafast across API, Codex and eligible ChatGPT Work plans at up to 8x Standard speed
3. FAR.AI’s preprint reports 98.8% AUC deception probes above its Opus 5.5 text-monitoring baseline on SHADE-Arena
4. Google previews its workplace Gemini agent for running sub-agents under their own identities
5. Hugging Face’s TRL v1.15.0 fuses the LM head for up to 6.9x longer Gemma-3-1B training sequences in single-GPU tests
6. Tencent releases 5B-parameter Youtu-Parsing-Omni to parse documents, images, audio and video into structured JSON
7. StepFun’s Step 5 Preview MoE reaches OpenRouter with 1M-token context and 27B active parameters
8. Arena previews its Alignment Index using 90,000 real-world agent sessions across 27 models
9. Philip Zeyliger shows how Shoulder lets AI agents watch and control an existing terminal
10. Geoffrey Huntley argues tool makers should improve CLI help before adding agent integrations
11. White House announces $2.4B in AI tools and compute credits for the Genesis Mission
12. GlobalFoundries signs a five-year, $2 billion TSMC deal for US-made AI-chip silicon interposers
13. Three fired OpenAI safety researchers deny misconduct in a newly published letter
14. USA Today sues OpenAI for more than $250 million over alleged AI-training copyright infringement
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
Launching an opt-in vulnerability-finding service for open-source software
We’re launching OSS Scanner, an opt-in vulnerability scanner for the open-source ecosystem informed by our experience using Claude to find vulnerabilities during Project Glasswing. Projects that join will receive thorough, periodic security scans by our strongest…
🗒 Calathea Daily AI Roundups · 10 Oct 2026
1. Anthropic cuts live internet access across internal evaluations after Claude bypassed restrictions and submitted real forms
2. Microsoft launches Microsoft-Decision-1 in Foundry, reporting 35x faster decisions than GPT-6 Sol in its tests
3. DatologyAI reports curated data matched its largest baseline model using 6x less pretraining compute
4. Thomas Hales argues Lean-checked AI proofs still need human checks that theorem statements match the intended mathematics
5. Simon Willison describes building a Django feature using about 30 minutes of voice prompts and 30 minutes of typed review
6. Nathan Lambert argues AI engineering will accelerate research without guaranteeing superhuman abilities beyond math and coding
7. Vercel CLI lets agents search, price and start domain purchases that require human approval
8. Google employees are testing unreleased Gemini 4 Carbon in its coding platform, Business Insider reports
9. TypeSafe AI raises $870 million at a $7.5 billion valuation for Jev in an a16z-led round
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic cuts live internet access across internal evaluations after Claude bypassed restrictions and submitted real forms
2. Microsoft launches Microsoft-Decision-1 in Foundry, reporting 35x faster decisions than GPT-6 Sol in its tests
3. DatologyAI reports curated data matched its largest baseline model using 6x less pretraining compute
4. Thomas Hales argues Lean-checked AI proofs still need human checks that theorem statements match the intended mathematics
5. Simon Willison describes building a Django feature using about 30 minutes of voice prompts and 30 minutes of typed review
6. Nathan Lambert argues AI engineering will accelerate research without guaranteeing superhuman abilities beyond math and coding
7. Vercel CLI lets agents search, price and start domain purchases that require human approval
8. Google employees are testing unreleased Gemini 4 Carbon in its coding platform, Business Insider reports
9. TypeSafe AI raises $870 million at a $7.5 billion valuation for Jev in an a16z-led round
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
Investigating unintended model actions in our evaluations and internal use
Anthropic is an AI safety and research company that's working to build reliable, interpretable, and steerable AI systems.
🗒 Calathea Daily AI Roundups · 11 Oct 2026
1. Microsoft CEO Satya Nadella calls for an AI emergency brake and deterministic controls against models’ insider risks.
2. Nish Tahir shows how to calibrate Qwen3-1.7B for single-pass probabilities instead of generated text.
3. BoxLite explains its plan to replace libkrun with a custom virtual machine monitor for AI-agent sandboxes.
4. Nvidia is in early talks to buy or invest more in open-weight model developer Reflection AI, the Financial Times reports.
5. BBC’s Ros Atkins examines why AI-generated culture may weaken the premium audiences place on human-made art.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Microsoft CEO Satya Nadella calls for an AI emergency brake and deterministic controls against models’ insider risks.
2. Nish Tahir shows how to calibrate Qwen3-1.7B for single-pass probabilities instead of generated text.
3. BoxLite explains its plan to replace libkrun with a custom virtual machine monitor for AI-agent sandboxes.
4. Nvidia is in early talks to buy or invest more in open-weight model developer Reflection AI, the Financial Times reports.
5. BBC’s Ros Atkins examines why AI-generated culture may weaken the premium audiences place on human-made art.
Follow Calathea for all things AI: https://t.me/CalatheaAI
sn scratchpad
Models as Insider Risks in the Super Intelligence Era
Thoughts on models as insider risks in the Super Intelligence Era.