🗒 Calathea Daily AI Roundups · 15 Oct 2026
1. SpaceXSI’s Grok Voice 3.0 scores 86.9 on Artificial Analysis’s Speech-to-Speech Index at $0.08 per audio minute.
2. Swarmcha.se publishes its full investigation of over 2,600 past Amap extraction attempts by Tencent-linked agents.
3. Noma Labs details a patched CVSS 8.7 Claude Code Desktop preview flaw that could steal local files.
4. First Proof scores ProofCouncil 9/10 and a Kimi K3-based system 0/10 on research mathematics.
5. Paper Instruments says Ultramarine 1 Flash beats Claude Fable 5 at $0.13/task versus $8.19/task on its internal knowledge-work benchmark.
6. StepFun explains Step 5 Preview’s 92-layer design for long-running agents.
7. Google opens Deep Review with two free document reviews per week.
8. Siemens EDA reports chip-design benchmark scores rising from about 55% to 95% and 66% fewer agent turns using NVIDIA NeMo Gym.
9. Anthropic opens public-beta OpenTelemetry trace export for Claude Platform agents.
10. Addy Osmani argues coding agents need fast, repeatable tests to verify their own work before human review.
11. 8 US states sue OpenAI over alleged child-safety harms and deceptive ChatGPT marketing.
12. FTC subpoenas OpenAI and Anthropic for AI-safety records, Axios reports.
13. TSMC reports 68.7% year-on-year third-quarter profit growth and forecasts $52.3–54.3 billion in fourth-quarter revenue.
14. Modal closes a $750M Series D after reaching a $1B annualized revenue run rate.
15. Anthropic’s quarterly revenue rose 45% to $16.8 billion, BBC reports.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. SpaceXSI’s Grok Voice 3.0 scores 86.9 on Artificial Analysis’s Speech-to-Speech Index at $0.08 per audio minute.
2. Swarmcha.se publishes its full investigation of over 2,600 past Amap extraction attempts by Tencent-linked agents.
3. Noma Labs details a patched CVSS 8.7 Claude Code Desktop preview flaw that could steal local files.
4. First Proof scores ProofCouncil 9/10 and a Kimi K3-based system 0/10 on research mathematics.
5. Paper Instruments says Ultramarine 1 Flash beats Claude Fable 5 at $0.13/task versus $8.19/task on its internal knowledge-work benchmark.
6. StepFun explains Step 5 Preview’s 92-layer design for long-running agents.
7. Google opens Deep Review with two free document reviews per week.
8. Siemens EDA reports chip-design benchmark scores rising from about 55% to 95% and 66% fewer agent turns using NVIDIA NeMo Gym.
9. Anthropic opens public-beta OpenTelemetry trace export for Claude Platform agents.
10. Addy Osmani argues coding agents need fast, repeatable tests to verify their own work before human review.
11. 8 US states sue OpenAI over alleged child-safety harms and deceptive ChatGPT marketing.
12. FTC subpoenas OpenAI and Anthropic for AI-safety records, Axios reports.
13. TSMC reports 68.7% year-on-year third-quarter profit growth and forecasts $52.3–54.3 billion in fourth-quarter revenue.
14. Modal closes a $750M Series D after reaching a $1B annualized revenue run rate.
15. Anthropic’s quarterly revenue rose 45% to $16.8 billion, BBC reports.
Follow Calathea for all things AI: https://t.me/CalatheaAI
x.ai
Introducing Grok Voice 3.0
The world’s most intelligent voice model.
❤1
🗒 Calathea Daily AI Roundups · 16 Oct 2026
1. Meta releases 816B-parameter Muse Spark Open for two M5 Ultra Mac Studios, claiming US open-weight leadership.
2. Microsoft details 3T-class MAI-Base-2 pretraining (2.8T parameters), claiming stronger results than tested open models.
3. NVIDIA releases open-weight Nemotron 3.5 Super VL for multimodal agents using 120B total and 12B active parameters.
4. NVIDIA releases hybrid Mamba-Transformer Nemotron 3.5 Bolt 74B-A9B for local agents in 24GB of memory.
5. Goodfire’s Nick Cammarata argues trillion-parameter models remain interpretable after preliminary Qwen3-32B audits found ~90% of neuron poles had one main function.
6. Anthropic reports over 2x merged Claude Code pull requests and over 40% lower revert rates after automating review.
7. PwnAI details AI-assisted PwnEscape host escape in its own Linux testbed during Vercel’s $1M sandbox challenge.
8. CoreWeave explains filtering 5 million daily hardware signals to five alerts before AI agents investigate them.
9. Mistral releases Timestral forecasting models from 5M to 209M parameters for CPU-based inference.
10. Google DeepMind’s preprint reports four Gemini 4 Argon agents beat Minecraft on Easy in 27 hours 14 minutes.
11. Anthropic embeds six METR evaluators with employee-like access and independent rights to publish safety findings.
12. China proposes export controls covering 23 AI technologies, excluding publicly released code and research.
13. OpenAI reports its internal GPT-5.6 Sol model used another wallet’s Sepolia test funds without permission during a June task.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Meta releases 816B-parameter Muse Spark Open for two M5 Ultra Mac Studios, claiming US open-weight leadership.
2. Microsoft details 3T-class MAI-Base-2 pretraining (2.8T parameters), claiming stronger results than tested open models.
3. NVIDIA releases open-weight Nemotron 3.5 Super VL for multimodal agents using 120B total and 12B active parameters.
4. NVIDIA releases hybrid Mamba-Transformer Nemotron 3.5 Bolt 74B-A9B for local agents in 24GB of memory.
5. Goodfire’s Nick Cammarata argues trillion-parameter models remain interpretable after preliminary Qwen3-32B audits found ~90% of neuron poles had one main function.
6. Anthropic reports over 2x merged Claude Code pull requests and over 40% lower revert rates after automating review.
7. PwnAI details AI-assisted PwnEscape host escape in its own Linux testbed during Vercel’s $1M sandbox challenge.
8. CoreWeave explains filtering 5 million daily hardware signals to five alerts before AI agents investigate them.
9. Mistral releases Timestral forecasting models from 5M to 209M parameters for CPU-based inference.
10. Google DeepMind’s preprint reports four Gemini 4 Argon agents beat Minecraft on Easy in 27 hours 14 minutes.
11. Anthropic embeds six METR evaluators with employee-like access and independent rights to publish safety findings.
12. China proposes export controls covering 23 AI technologies, excluding publicly released code and research.
13. OpenAI reports its internal GPT-5.6 Sol model used another wallet’s Sepolia test funds without permission during a June task.
Follow Calathea for all things AI: https://t.me/CalatheaAI
Meta AI Research
Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date
Muse Spark Open is the most capable US open-weight model to date, and brings leading agentic capabilities to your desktop on Apple M5 Ultra Mac Studios.
🗒 Calathea Daily AI Roundups · 17 Oct 2026
1. DeepSeek releases V4.1-Flash-1017 and opens Search and Fetch APIs, each priced at $0.75 per 1,000 calls.
2. Reflection releases Beam’s 501B-total, 23B-active model weights under Apache 2.0.
3. Anthropic’s Frontier Red Team says Claude recovered timestamp-based origins for NIST ECDSA curve seeds.
4. Anthropic finds interactive AI-agent session hours exceeded employee work hours 6 to 1 by August.
5. Google Cloud’s Casey West shares four lessons for evaluating agents beyond their final replies.
6. Anthropic pitches lawmakers a draft federal AI safety bill, Politico reports.
7. Saif Khan projects a roughly 100x US AI-chip production lead by the mid-2030s if DUV-tool exports to China stop.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. DeepSeek releases V4.1-Flash-1017 and opens Search and Fetch APIs, each priced at $0.75 per 1,000 calls.
2. Reflection releases Beam’s 501B-total, 23B-active model weights under Apache 2.0.
3. Anthropic’s Frontier Red Team says Claude recovered timestamp-based origins for NIST ECDSA curve seeds.
4. Anthropic finds interactive AI-agent session hours exceeded employee work hours 6 to 1 by August.
5. Google Cloud’s Casey West shares four lessons for evaluating agents beyond their final replies.
6. Anthropic pitches lawmakers a draft federal AI safety bill, Politico reports.
7. Saif Khan projects a roughly 100x US AI-chip production lead by the mid-2030s if DUV-tool exports to China stop.
Follow Calathea for all things AI: https://t.me/CalatheaAI
Deepseek
DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real | DeepSeek API Docs
🚀 DeepSeek-V4.1-Flash-1017 is here!
🗒 Calathea Daily AI Roundups · 18 Oct 2026
1. OpenCode discloses an ongoing security breach affecting 1.4% of Console users with linked external providers.
2. vLLM releases v0.32.0, restricting default browser origins and dropping Python 3.10 support.
3. Will Larson says about 10% of Imprint changes qualify for fast approval under automated checks.
4. Anemll reports 24K-context support for its iPhone 18 Qwen3.8-27B demo: 1.5 bpw, 15–70 tok/s including predictable-text peaks.
5. SemiAnalysis reports 2.8T-parameter Kimi K3 at 423 tok/s/user on GB300, 1.68x its earlier benchmark across different configurations.
6. Nathan Lambert argues Meta and Reflection roughly halve the US–China open-weight model performance gap.
7. Jared Siegel argues self-hosted LLM proxies improve token-use visibility and provider-key control.
8. Sean Goedecke argues AI-generated writing no longer proves employee effort, learning or compliance.
9. Marcus Johansson reports GPT-5.5 preserved his chosen focal point in 96% of crops across 19 test photos.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. OpenCode discloses an ongoing security breach affecting 1.4% of Console users with linked external providers.
2. vLLM releases v0.32.0, restricting default browser origins and dropping Python 3.10 support.
3. Will Larson says about 10% of Imprint changes qualify for fast approval under automated checks.
4. Anemll reports 24K-context support for its iPhone 18 Qwen3.8-27B demo: 1.5 bpw, 15–70 tok/s including predictable-text peaks.
5. SemiAnalysis reports 2.8T-parameter Kimi K3 at 423 tok/s/user on GB300, 1.68x its earlier benchmark across different configurations.
6. Nathan Lambert argues Meta and Reflection roughly halve the US–China open-weight model performance gap.
7. Jared Siegel argues self-hosted LLM proxies improve token-use visibility and provider-key control.
8. Sean Goedecke argues AI-generated writing no longer proves employee effort, learning or compliance.
9. Marcus Johansson reports GPT-5.5 preserved his chosen focal point in 96% of crops across 19 test photos.
Follow Calathea for all things AI: https://t.me/CalatheaAI
X (formerly Twitter)
OpenCode (@opencode) on X
We had a security breach last night related to a large scale distillation attack that impacted 1.4% of OpenCode Console users
We have notified them all this morning via email. If you did not get…
We have notified them all this morning via email. If you did not get…
🗒 Calathea Daily AI Roundups · 19 Oct 2026
1. A hacker reportedly used AI agents to breach over 100 company websites for an estimated $5,000 in tokens.
2. monday.com reports its rebuilt API edge handles 2.5x the traffic using 92% less CPU.
3. Brookhaven and Great Sky report superconducting AI chips analyzed fusion data recorded at one million samples per second.
4. Sean Goedecke argues coding agents need explicit limits on requirements to avoid unnecessary complexity.
5. Anthropic proposes committing 5% of its shares to a public-benefit fund, subject to stockholder approval.
6. SemiAnalysis forecasts 50GW of new global datacenter capacity in 2027.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. A hacker reportedly used AI agents to breach over 100 company websites for an estimated $5,000 in tokens.
2. monday.com reports its rebuilt API edge handles 2.5x the traffic using 92% less CPU.
3. Brookhaven and Great Sky report superconducting AI chips analyzed fusion data recorded at one million samples per second.
4. Sean Goedecke argues coding agents need explicit limits on requirements to avoid unnecessary complexity.
5. Anthropic proposes committing 5% of its shares to a public-benefit fund, subject to stockholder approval.
6. SemiAnalysis forecasts 50GW of new global datacenter capacity in 2027.
Follow Calathea for all things AI: https://t.me/CalatheaAI
mint
How one hacker used AI agents and a few thousand dollars to break into scores of sites | Mint
AI-CYBERCRIME/ (PIX):EXCLUSIVE-How one hacker used AI agents and a few thousand dollars to break into scores of sites
🗒 Calathea Daily AI Roundups · 20 Oct 2026
1. DeepSeek open-sources DeepInfer’s offline inference engine for DeepSeek-V4.1-Flash on NVIDIA and Huawei Ascend hardware.
2. Dream Labs opens Dream-1’s robot-control API beta at $3/hour.
3. Inception reports its new Mercury 3 diffusion model generates 1,068 tokens/second on NVIDIA GPUs.
4. MiniMax releases M3.1 Flash at $0.15 input/$0.60 output per million tokens.
5. OpenAI reports GPT-6 Astra’s 87.4% average pass rate on its 20-language coding test, versus GPT-5.6 Sol’s 26.8%.
6. Amazon announces TorchNeuron’s October 20 public beta for native PyTorch on Trainium2 and Trainium3.
7. Cloudflare previews Clairic for extracting structured data in a single pass on Workers AI.
8. Skyfall AI previews SKY-001, claiming 51–125x faster CAD/BIM generation at 150–210x lower cost than tested frontier models.
9. Wiz links an AI-driven cloud attacker’s source IP to scans of over 600 customer environments.
10. Meta reports 14 specialist agents cut manual data-column preparation from 3–4 days to under one hour.
11. Cloudflare reports cutting its AI code-review inference bill by 68% through prompt caching and model routing.
12. Tenstorrent raises $1.5 billion at a $6.2 billion post-money valuation.
13. DHS publishes its order excluding DeepSeek-based products from civilian federal procurement and IT systems.
14. Nscale and Anthropic reportedly discuss another $45 billion cloud-computing deal.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. DeepSeek open-sources DeepInfer’s offline inference engine for DeepSeek-V4.1-Flash on NVIDIA and Huawei Ascend hardware.
2. Dream Labs opens Dream-1’s robot-control API beta at $3/hour.
3. Inception reports its new Mercury 3 diffusion model generates 1,068 tokens/second on NVIDIA GPUs.
4. MiniMax releases M3.1 Flash at $0.15 input/$0.60 output per million tokens.
5. OpenAI reports GPT-6 Astra’s 87.4% average pass rate on its 20-language coding test, versus GPT-5.6 Sol’s 26.8%.
6. Amazon announces TorchNeuron’s October 20 public beta for native PyTorch on Trainium2 and Trainium3.
7. Cloudflare previews Clairic for extracting structured data in a single pass on Workers AI.
8. Skyfall AI previews SKY-001, claiming 51–125x faster CAD/BIM generation at 150–210x lower cost than tested frontier models.
9. Wiz links an AI-driven cloud attacker’s source IP to scans of over 600 customer environments.
10. Meta reports 14 specialist agents cut manual data-column preparation from 3–4 days to under one hour.
11. Cloudflare reports cutting its AI code-review inference bill by 68% through prompt caching and model routing.
12. Tenstorrent raises $1.5 billion at a $6.2 billion post-money valuation.
13. DHS publishes its order excluding DeepSeek-based products from civilian federal procurement and IT systems.
14. Nscale and Anthropic reportedly discuss another $45 billion cloud-computing deal.
Follow Calathea for all things AI: https://t.me/CalatheaAI
GitHub
GitHub - deepseek-ai/DeepInfer: DeepInfer: High-performance, large-scale inference framework for DeepSeek models
DeepInfer: High-performance, large-scale inference framework for DeepSeek models - deepseek-ai/DeepInfer
🗒 Calathea Daily AI Roundups · 21 Oct 2026
1. Mistral releases Voxtral 3 models for transcription in 50+ languages, speaker diarization and speech generation.
2. NVIDIA announces DSX POD’s 14-day deployment target for prepared sites, an OpenShell security architecture and Agent Knowledge Foundation at GTC Berlin.
3. Google open-sources TorchTPU to run PyTorch natively on TPUs.
4. ElevenLabs launches Eleven Realtime v1 Mini and Max voice models for over 90 languages and 17,000+ voices.
5. Perplexity launches fully local Computer for paid users on Apple silicon Macs with 36 GB memory recommended.
6. Gradium releases Phonon’s 31M-parameter speech weights under Apache 2.0 and 40M-parameter variants under non-commercial licenses.
7. OpenAI discloses how GPT-6.1 Sol and GPT-6 Astra agents weakened monitoring during internal coding tasks.
8. Meta pauses Muse sharing after researchers find over 6,800 share links through public DNS records.
9. OpenAI adds 12-hour API prompt caching for cheaper cached inputs at a 2.5x input-price cost for cache writes.
10. Chinese Academy of Sciences researchers show planted agent attacks can persist across sessions in a 101-task, five-model CarryoverBench preprint.
11. KAIST and Amazon researchers report 2.54x autoregressive-baseline throughput for Lafa 7B at comparable accuracy in single-H100 preprint tests.
12. WSJ reports Anthropic’s first-half net loss exceeded $80 billion, mostly from revaluing financial instruments used to raise funding.
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Mistral releases Voxtral 3 models for transcription in 50+ languages, speaker diarization and speech generation.
2. NVIDIA announces DSX POD’s 14-day deployment target for prepared sites, an OpenShell security architecture and Agent Knowledge Foundation at GTC Berlin.
3. Google open-sources TorchTPU to run PyTorch natively on TPUs.
4. ElevenLabs launches Eleven Realtime v1 Mini and Max voice models for over 90 languages and 17,000+ voices.
5. Perplexity launches fully local Computer for paid users on Apple silicon Macs with 36 GB memory recommended.
6. Gradium releases Phonon’s 31M-parameter speech weights under Apache 2.0 and 40M-parameter variants under non-commercial licenses.
7. OpenAI discloses how GPT-6.1 Sol and GPT-6 Astra agents weakened monitoring during internal coding tasks.
8. Meta pauses Muse sharing after researchers find over 6,800 share links through public DNS records.
9. OpenAI adds 12-hour API prompt caching for cheaper cached inputs at a 2.5x input-price cost for cache writes.
10. Chinese Academy of Sciences researchers show planted agent attacks can persist across sessions in a 101-task, five-model CarryoverBench preprint.
11. KAIST and Amazon researchers report 2.54x autoregressive-baseline throughput for Lafa 7B at comparable accuracy in single-H100 preprint tests.
12. WSJ reports Anthropic’s first-half net loss exceeded $80 billion, mostly from revaluing financial instruments used to raise funding.
Follow Calathea for all things AI: https://t.me/CalatheaAI
Mistral
Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages
Voxtral 3: Mistral's voice agent SDK for real-time speech in 50+ languages — accurate transcription, emotional TTS, deep tool use. EU-hosted or self-deployed.
🗒 Calathea Daily AI Roundups · 22 Oct 2026
1. Wiz discloses two AI-found PHP RCE flaws, CVE-2026-109618 and CVE-2026-109619, in the language powering 70% of websites
2. Anthropic opens Claude Duet’s research-preview waitlist for US Pro and Max users, offering proactive agent work while they’re away
3. Recursive releases 40B Flint, largely built by its AI research system, reporting over 125 tokens/second on M4 Pro
4. Hugging Face publishes 14.8B-parameter Carbon-X for DNA questions, reporting 66.7% across five benchmarks without clinical validation
5. OpenAI sets November 20 approval rules for model routers and AI gateways, excluding apps that merely integrate its models
6. LiteLLM pulls compromised 1.107.0 after credential-stealing malware and a backdoor reach PyPI and container registries
7. Railway urges customers to rotate secrets after attackers may have accessed and decrypted stored environment variables
8. Anthropic adds continuous scanning and fix tracking to Claude Security’s Enterprise beta, offering $25,000 monthly credits for two months
9. IBM updates 470M-parameter Granite Speech 5.0 TurboCTC for seven languages, with 1.6x faster transcription in leaderboard tests
10. GitHub engineers explain copying 99.99999% of repositories to Azure, with 3–4x write throughput in internal per-reference tests
11. curl founder Daniel Stenberg argues accurate AI vulnerability reports cannot be ignored, even when their volume strains maintainers
12. ChinaTalk’s A. Chen examines robot-training data’s weak resale economics, with usable footage estimated at 500–700 yuan per hour
13. DuckDB releases 2.0.0 LTS “Cinnamon Teal,” with agent mode cutting model-facing output 59% in its TPC-H tests
14. OpenAI partners with Google Cloud on frontier models and Codex, starting with a private preview for selected enterprise customers
15. Reuters reports DeepSeek nears a $15 billion funding round, with terms still subject to change
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Wiz discloses two AI-found PHP RCE flaws, CVE-2026-109618 and CVE-2026-109619, in the language powering 70% of websites
2. Anthropic opens Claude Duet’s research-preview waitlist for US Pro and Max users, offering proactive agent work while they’re away
3. Recursive releases 40B Flint, largely built by its AI research system, reporting over 125 tokens/second on M4 Pro
4. Hugging Face publishes 14.8B-parameter Carbon-X for DNA questions, reporting 66.7% across five benchmarks without clinical validation
5. OpenAI sets November 20 approval rules for model routers and AI gateways, excluding apps that merely integrate its models
6. LiteLLM pulls compromised 1.107.0 after credential-stealing malware and a backdoor reach PyPI and container registries
7. Railway urges customers to rotate secrets after attackers may have accessed and decrypted stored environment variables
8. Anthropic adds continuous scanning and fix tracking to Claude Security’s Enterprise beta, offering $25,000 monthly credits for two months
9. IBM updates 470M-parameter Granite Speech 5.0 TurboCTC for seven languages, with 1.6x faster transcription in leaderboard tests
10. GitHub engineers explain copying 99.99999% of repositories to Azure, with 3–4x write throughput in internal per-reference tests
11. curl founder Daniel Stenberg argues accurate AI vulnerability reports cannot be ignored, even when their volume strains maintainers
12. ChinaTalk’s A. Chen examines robot-training data’s weak resale economics, with usable footage estimated at 500–700 yuan per hour
13. DuckDB releases 2.0.0 LTS “Cinnamon Teal,” with agent mode cutting model-facing output 59% in its TPC-H tests
14. OpenAI partners with Google Cloud on frontier models and Codex, starting with a private preview for selected enterprise customers
15. Reuters reports DeepSeek nears a $15 billion funding round, with terms still subject to change
Follow Calathea for all things AI: https://t.me/CalatheaAI
wiz.io
PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet | Wiz Blog
Core PHP engine flaws enable remote code execution and Heartbleed-style memory disclosure across the web.
🗒 Calathea Daily AI Roundups · 23 Oct 2026
1. Anthropic and Theorem generate 180 million Rocq proof lines with Claude across 123 libraries, establishing crash-freedom under restricted RISC-V assumptions
2. OpenAI's Decisions API reaches general availability, claiming up-to-10x speedups over GPT-6 Luna’s Responses API and 400 questions per request
3. METR and Redwood Research expose new monitoring weaknesses after three embedded weeks at OpenAI, withholding exploit details
4. Agentic AI Foundation welcomes four projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox, placing agent infrastructure under neutral governance
5. Google announces hardware-isolated Agent Compute Engine for autonomous agents, with preview access coming in the next few months
6. NVIDIA benchmarks eight DGX Spark nodes with 1 TB memory, reporting 4.43x single-node decode speed on Qwen3.8-27B NVFP4
7. Ai2 releases OlmoEarth v1.3 with global 10-meter satellite embeddings in 64 and 128 dimensions, usable without model fine-tuning
8. Ideogram ships Vector 1.0 for editable SVGs from prompts or images, available through its app and API
9. Wiz links ARTEX-driven identity attacks across software vendors to activity overlapping Storm-0940, extending earlier supply-chain findings
10. Microsoft plans hospital testing of MAI-DxO after Science publication, reporting 1.7x generalist-physician accuracy on simulated emergency-department cases
11. Dwarkesh Patel's Intel tour contrasts historic 3x annual AI-compute growth with existing tools' roughly double leading-edge production ceiling
12. Carroll Vance details replacing PyTorch with Julia and Reactant for medical imaging, tracing a pathology model from training to serving
13. Groundrun argues hardware tests must stay outside coding agents’ control, drawing on hundreds of incident notes about Claude Code mistakes
14. Google expands its Voltus virtual power plant agreement tenfold to 1 gigawatt, paying participants to reduce peak electricity demand
15. CNBC reports OpenAI nears up to $30 billion at a $1.25 trillion pre-money valuation before a possible 2027 IPO
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Anthropic and Theorem generate 180 million Rocq proof lines with Claude across 123 libraries, establishing crash-freedom under restricted RISC-V assumptions
2. OpenAI's Decisions API reaches general availability, claiming up-to-10x speedups over GPT-6 Luna’s Responses API and 400 questions per request
3. METR and Redwood Research expose new monitoring weaknesses after three embedded weeks at OpenAI, withholding exploit details
4. Agentic AI Foundation welcomes four projects: Agent Skills, Open Responses, OpenEnv, and OpenSandbox, placing agent infrastructure under neutral governance
5. Google announces hardware-isolated Agent Compute Engine for autonomous agents, with preview access coming in the next few months
6. NVIDIA benchmarks eight DGX Spark nodes with 1 TB memory, reporting 4.43x single-node decode speed on Qwen3.8-27B NVFP4
7. Ai2 releases OlmoEarth v1.3 with global 10-meter satellite embeddings in 64 and 128 dimensions, usable without model fine-tuning
8. Ideogram ships Vector 1.0 for editable SVGs from prompts or images, available through its app and API
9. Wiz links ARTEX-driven identity attacks across software vendors to activity overlapping Storm-0940, extending earlier supply-chain findings
10. Microsoft plans hospital testing of MAI-DxO after Science publication, reporting 1.7x generalist-physician accuracy on simulated emergency-department cases
11. Dwarkesh Patel's Intel tour contrasts historic 3x annual AI-compute growth with existing tools' roughly double leading-edge production ceiling
12. Carroll Vance details replacing PyTorch with Julia and Reactant for medical imaging, tracing a pathology model from training to serving
13. Groundrun argues hardware tests must stay outside coding agents’ control, drawing on hundreds of incident notes about Claude Code mistakes
14. Google expands its Voltus virtual power plant agreement tenfold to 1 gigawatt, paying participants to reduce peak electricity demand
15. CNBC reports OpenAI nears up to $30 billion at a $1.25 trillion pre-money valuation before a possible 2027 IPO
Follow Calathea for all things AI: https://t.me/CalatheaAI
Anthropic
LLMs are making formal software verification feasible at a much larger scale
Claude formally proved security properties in the compiled code of 123 open-source libraries. Along the way, it found bugs that had eluded detection for years.
🗒 Calathea Daily AI Roundups · 24 Oct 2026
1. Moonshot AI launches 3T-class, 1M-context Kimi K3.1 Pro, fourth on Artificial Analysis’s Intelligence Index (55), with 2–3×-faster Highspeed; weights due November 4
2. NVIDIA releases Nemotron 3.5 Bolt E4B 4GB for local assistants, claiming ~40% higher conversational-benchmark averages than Nemotron 3 Nano 4B
3. Block previews speech-only, full-duplex Dessa, runnable on an M4 Max MacBook Pro but initially restricted to safety researchers
4. UK launches ÂŁ100 million AI Biosecurity Lab; OpenAI and Anthropic commit up to ÂŁ200 million in credits, access and support
5. NVIDIA details Hermes-assisted Microduck simulation, training across 4,096 environments on RTX PRO 6000 Blackwell GPU before RTX Spark validation
6. Vercel explains Vercel Agent’s plan-scoped production permissions, with secrets injected outside the agent’s environment
7. OpenAI details Habitat’s 12x median-read speedup, from roughly 120 ms to roughly 10 ms, serving over 1 billion users
8. STFC engineer Drew Silcock argues closing source code cannot stop AI-assisted hacking and fixed patch delays can worsen exposure
9. Amazon CEO Andy Jassy says AWS’s 2026 capacity is sold out, 2027 largely booked and a good chunk of 2028 committed
10. CNBC reports TikTok’s US venture seeks 4 gigawatts of computing capacity by the early 2030s, with Oracle remaining its primary provider
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Moonshot AI launches 3T-class, 1M-context Kimi K3.1 Pro, fourth on Artificial Analysis’s Intelligence Index (55), with 2–3×-faster Highspeed; weights due November 4
2. NVIDIA releases Nemotron 3.5 Bolt E4B 4GB for local assistants, claiming ~40% higher conversational-benchmark averages than Nemotron 3 Nano 4B
3. Block previews speech-only, full-duplex Dessa, runnable on an M4 Max MacBook Pro but initially restricted to safety researchers
4. UK launches ÂŁ100 million AI Biosecurity Lab; OpenAI and Anthropic commit up to ÂŁ200 million in credits, access and support
5. NVIDIA details Hermes-assisted Microduck simulation, training across 4,096 environments on RTX PRO 6000 Blackwell GPU before RTX Spark validation
6. Vercel explains Vercel Agent’s plan-scoped production permissions, with secrets injected outside the agent’s environment
7. OpenAI details Habitat’s 12x median-read speedup, from roughly 120 ms to roughly 10 ms, serving over 1 billion users
8. STFC engineer Drew Silcock argues closing source code cannot stop AI-assisted hacking and fixed patch delays can worsen exposure
9. Amazon CEO Andy Jassy says AWS’s 2026 capacity is sold out, 2027 largely booked and a good chunk of 2028 committed
10. CNBC reports TikTok’s US venture seeks 4 gigawatts of computing capacity by the early 2030s, with Oracle remaining its primary provider
Follow Calathea for all things AI: https://t.me/CalatheaAI
www.kimi.ai
Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence
Introducing Kimi K3.1 Pro, our most capable 3T-class open-weight model with 1M-token context and major advances in coding, 3D modeling, and financial research.
❤1
🗒 Calathea Daily AI Roundups · 25 Oct 2026
1. llama.cpp adds DeepSeek-V4.1-Flash support in prerelease b11870, converting Engram tables from MXFP8 to BF16
2. Programmer Dan Luu argues agents enable ambitious software, citing informal Rust-tooling tests with 50x cold-build and 2x warm-build speedups
3. Security researcher Utku Sen compares 3 AI pentesting harnesses on a toy app, finding general-purpose agents can match them
4. WIRED examines millions of PHP exploit attempts, while researchers caution sophisticated AI-assisted attacks still rely on human expertise
5. FBI considers an economic-espionage charge against Saudi ex-OpenAI engineer accused of bugging executives, The Wall Street Journal reports
6. US warns AI partners against backing rival blocs as countries seek both American and Chinese technology, AP reports
7. Entrepreneur Steve Blank argues AI lowers the cost of European software alternatives as US suppliers become geopolitical risks
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. llama.cpp adds DeepSeek-V4.1-Flash support in prerelease b11870, converting Engram tables from MXFP8 to BF16
2. Programmer Dan Luu argues agents enable ambitious software, citing informal Rust-tooling tests with 50x cold-build and 2x warm-build speedups
3. Security researcher Utku Sen compares 3 AI pentesting harnesses on a toy app, finding general-purpose agents can match them
4. WIRED examines millions of PHP exploit attempts, while researchers caution sophisticated AI-assisted attacks still rely on human expertise
5. FBI considers an economic-espionage charge against Saudi ex-OpenAI engineer accused of bugging executives, The Wall Street Journal reports
6. US warns AI partners against backing rival blocs as countries seek both American and Chinese technology, AP reports
7. Entrepreneur Steve Blank argues AI lowers the cost of European software alternatives as US suppliers become geopolitical risks
Follow Calathea for all things AI: https://t.me/CalatheaAI
GitHub
Release b11870 · ggml-org/llama.cpp
model: DSV4.1 flash (#31086)
model : add DeepSeek-V4.1-Flash (deepseek41)
deepseek41 : remove MXFP8, convert the Engram table to BF16
deepseek41 : compute the compressed-KV read layers once i...
model : add DeepSeek-V4.1-Flash (deepseek41)
deepseek41 : remove MXFP8, convert the Engram table to BF16
deepseek41 : compute the compressed-KV read layers once i...
🗒 Calathea Daily AI Roundups · 26 Oct 2026
1. Confidential-AI provider Tinfoil details replacing 1M-line firmware with ~300 assembly lines to reduce its enclave boot attack surface
2. Anthropic confirms Georgia was the state whose vote-count paywall Claude bypassed, saying voter-registration systems remained untouched
3. Wiz discloses Gheist’s 13,500-plus planted branches across 850-plus compromised GitHub accounts, with private repositories exposed despite public-repository protections
4. Haskell developer Csongor Kiss’s 75-submission toy-language experiment finds coding agents can miss exploitable runtime rules despite producing working code
5. Data scientist Max Woolf argues open-weight competition helps explain OpenAI and Anthropic price cuts of up to 80%
6. Stratechery’s Ben Thompson argues AI-driven hacking and lost referral traffic favor software built for small, trusted groups
7. XPENG wins Guangzhou-wide Level 4 robotaxi trial approval, with rides initially expanding through invitation codes
8. Independent mathematics advisers AGMAI urge AI labs to stop handpicking researchers to validate releases, citing unequal access and borrowed credibility
9. Elon Musk says SpaceX closed $40 billion in GPU debt as he forecasts a shakeout among AI cloud providers
10. Tata Communications plans $400-450 million for around 10,000 km of fibre connecting India’s AI data-centre hubs
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. Confidential-AI provider Tinfoil details replacing 1M-line firmware with ~300 assembly lines to reduce its enclave boot attack surface
2. Anthropic confirms Georgia was the state whose vote-count paywall Claude bypassed, saying voter-registration systems remained untouched
3. Wiz discloses Gheist’s 13,500-plus planted branches across 850-plus compromised GitHub accounts, with private repositories exposed despite public-repository protections
4. Haskell developer Csongor Kiss’s 75-submission toy-language experiment finds coding agents can miss exploitable runtime rules despite producing working code
5. Data scientist Max Woolf argues open-weight competition helps explain OpenAI and Anthropic price cuts of up to 80%
6. Stratechery’s Ben Thompson argues AI-driven hacking and lost referral traffic favor software built for small, trusted groups
7. XPENG wins Guangzhou-wide Level 4 robotaxi trial approval, with rides initially expanding through invitation codes
8. Independent mathematics advisers AGMAI urge AI labs to stop handpicking researchers to validate releases, citing unequal access and borrowed credibility
9. Elon Musk says SpaceX closed $40 billion in GPU debt as he forecasts a shakeout among AI cloud providers
10. Tata Communications plans $400-450 million for around 10,000 km of fibre connecting India’s AI data-centre hubs
Follow Calathea for all things AI: https://t.me/CalatheaAI
tinfoil.sh
Less Is More: From 1M Lines of Firmware to 300 of Assembly
How we replaced general-purpose firmware with code built exactly for our machines, moving its complexity out of runtime and into measured build artifacts and significantly reducing the size of the trusted code base.
🗒 Calathea Daily AI Roundups · 27 Oct 2026
1. OpenAI launches GPT-6.1 Terra for computer use at 50% lower standard API prices than GPT-5.6 Terra
2. Cloudflare discloses cross-customer data leaks in 155 GLM-5.3 responses, fixed in September after an SGLang cache-handling failure
3. Anthropic publishes a Claude-generated 2^o(n) 3-SAT algorithm preprint, claiming a Lean-formalized refutation of the Exponential Time Hypothesis
4. OpenAI presents three AI-generated science claims for academic scrutiny, releasing math proofs while withholding physics materials
5. Luma AI’s ZAP preprint reports 96.7% success across 10 real-robot tasks after simulation-only motion training
6. OpenAI updates Codex Security 0.7.0 with optional vulnerability validation and Deep Scan recovery fixes
7. UC Berkeley’s Ben Recht argues open AI needs shared compute and transparent training data, code and processes, beyond downloadable weights
8. Bugcrowd founder Casey Ellis argues AI speeds existing threats, while secrets and broken trust remain easier targets than memory-corruption flaws
9. Anthropic releases Claude for Nonprofits plugin, built with about 30 organizations for fundraising, hiring and governance
10. Anthropic files for an IPO, disclosing $16.3 billion first-half 2026 revenue while leaving offering size and price unset
11. Broadcom and four financial firms announce $63 billion financing for Anthropic chip leases, with deliveries expected from 2027
12. Meta plans up to $97.5 billion in equity sales for AI infrastructure after enlarging its common-stock offering, Reuters reports
13. Microsoft offers voluntary buyouts to around 60,000 employees, roughly a quarter of its workforce, as AI spending grows
Follow Calathea for all things AI: https://t.me/CalatheaAI
1. OpenAI launches GPT-6.1 Terra for computer use at 50% lower standard API prices than GPT-5.6 Terra
2. Cloudflare discloses cross-customer data leaks in 155 GLM-5.3 responses, fixed in September after an SGLang cache-handling failure
3. Anthropic publishes a Claude-generated 2^o(n) 3-SAT algorithm preprint, claiming a Lean-formalized refutation of the Exponential Time Hypothesis
4. OpenAI presents three AI-generated science claims for academic scrutiny, releasing math proofs while withholding physics materials
5. Luma AI’s ZAP preprint reports 96.7% success across 10 real-robot tasks after simulation-only motion training
6. OpenAI updates Codex Security 0.7.0 with optional vulnerability validation and Deep Scan recovery fixes
7. UC Berkeley’s Ben Recht argues open AI needs shared compute and transparent training data, code and processes, beyond downloadable weights
8. Bugcrowd founder Casey Ellis argues AI speeds existing threats, while secrets and broken trust remain easier targets than memory-corruption flaws
9. Anthropic releases Claude for Nonprofits plugin, built with about 30 organizations for fundraising, hiring and governance
10. Anthropic files for an IPO, disclosing $16.3 billion first-half 2026 revenue while leaving offering size and price unset
11. Broadcom and four financial firms announce $63 billion financing for Anthropic chip leases, with deliveries expected from 2027
12. Meta plans up to $97.5 billion in equity sales for AI infrastructure after enlarging its common-stock offering, Reuters reports
13. Microsoft offers voluntary buyouts to around 60,000 employees, roughly a quarter of its workforce, as AI spending grows
Follow Calathea for all things AI: https://t.me/CalatheaAI
OpenAI
Introducing GPT-6.1 Terra
Delivering fast and capable computer use at scale