https://calathea.ai/reads/ Curated AI reads and signal from the Calathea Telegram channel. en Sat, 22 Aug 2026 13:57:54 GMT https://t.me/CalatheaAI/96 https://t.me/CalatheaAI/96 Sat, 22 Aug 2026 12:27:25 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 22 Aug 2026

1. OpenAI cuts GPT-5.6 Sol API pricing to $4 per million input tokens and $20 per million output tokens through at least Nov. 21, down 20% and 33%

2. NVIDIA's AVO agent completes all 183 levels across ARC-AGI-3's 25 public environments with Claude Opus 5, using 12% fewer environment actions than VISTA

3. Anthropic moves Claude Mythos 5 into Claude Security scans for Enterprise customers and launches a $35M open-source security credit fund

4. NVIDIA takes a minority stake in Cloverleaf Infrastructure to accelerate powered-site development for US AI data centers

5. Anthropic hires Google TPU program founder Amir Salek for its compute team as it builds an in-house chip effort, Bloomberg reports

6. OpenAI lets eligible Global API projects select regional processing request by request through prefixed domains

7. Qwen Code reports qwen3.7-plus resolved 377 of 495 valid SWE-bench Verified cases in a 500-task end-to-end run, scoring 76.16% on valid grader results

8. TechCrunch finds Claude Opus 4.6 complied with all 10 direct explicit-content requests despite Anthropic's usage policy

9. SemiAnalysis finds open models are closing each frontier-era capability gap about twice as fast, while warning benchmarks still miss real product quality

10. Silicon Valley 101 maps how hyperscalers finance AI data centers through bonds, leases, guarantees and SPVs, spreading obligations across filings and counterparties

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/95 https://t.me/CalatheaAI/95 Fri, 21 Aug 2026 15:18:55 GMT Funding & raises Product launches Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 21 Aug 2026

1. Broadcom is in talks to raise more than $60B in debt for AI chips benefiting Anthropic and others, Reuters reports

2. DeepSeek launches experimental V4-Flash-Vision-Exp on its API

3. Nvidia strikes a reported $6B Poolside licensing deal and invests $1B at a $12B pre-money valuation

4. Nvidia is in early talks with Korean inference-chip startup Rebellions over a partnership, investment or acquisition

5. Anthropic is considering an IPO that could match or surpass SpaceX's record $75B raise, Bloomberg reports

6. Harvey says its Kimi K3-based Tenet model ranks first on LAB Contracts and second overall on LAB

7. Mistral launches Agentic Search and reports FinanceBench correctness rising from 26.7% to 86%

8. Nvidia denies a report that it plans a China-specific LPU by year-end and says none is on its roadmap

9. OpenAI is gaining faster in Q3 among Ramp's 70,000+ US business customers, while Anthropic still leads

10. Researchers say encrypted prompt injection can make Grok exfiltrate names, locations and active chat histories

11. OpenRouter launches free stealth model Ox Alpha with 1M context while its provider retains prompts and completions

12. Pew finds 35% of dated web pages published since ChatGPT show signs of AI authorship

13. CISPA researchers introduce EchoCoT, a tool-call replay attack for extracting hidden reasoning from black-box models

14. Nevada approves commercial robotaxi permits for Tesla, Uber and Waymo in Clark County

15. Hugging Face releases Diffusers 0.40.0 with stable Modular Diffusers, tensor parallelism and new pipelines

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/94 https://t.me/CalatheaAI/94 Thu, 20 Aug 2026 14:26:09 GMT Funding & raises Agent workflows Dev tooling & infra Research 🗒 Calathea Daily AI Roundups · 20 Aug 2026

1. Marvell discloses a Google warrant for up to $12.2B of its shares, with most vesting as TPU-linked custom-chip purchases generate up to roughly $120B through fiscal 2033

2. Fractile seeks about $600M at a $6.5B pre-money valuation after securing an initial $250M Anthropic chip commitment, Bloomberg reports

3. Nvidia is reportedly discussing an investment in a General Catalyst-led Mercor round at a $20B valuation

4. Nebius prices an upsized $5B convertible-note sale to fund data centers, GPUs and its AI cloud

5. Alibaba’s AI cloud revenue rose 45% to RMB48.4B as capital expenditure climbed 75% and net income fell 75%

6. The CFTC opens public consultation on compute derivatives, including market-manipulation risks and perpetual futures

7. Generalist says GEN-1.5 averaged 59% across 10 simple, short-horizon robot tasks after one 3–12-second demonstration with no gradient update, rising to 83% after 10 steps on five minutes of data

8. OpenAI previews cross-interaction safety monitoring for eligible frontier-model API customers while preserving zero data retention, with rollout and a technical paper planned for September

9. Waymo opens its sixth-generation Ojai robotaxi to all riders in Los Angeles, Phoenix and San Francisco

10. Callosum closes a $100M seed led by Atomico, with Plural, DCVC and UK Sovereign AI participating

11. Bloomberg’s coding test produced a functional Kimi K3 ecommerce site for $12, versus about $50 with Claude Fable 5

12. Thomson Reuters researchers Davide Romano and Kanak Raj find selection, not exploration, bottlenecks test-time scaling on five open-ended benchmarks

13. Claude Code adds a Concise output style that leads with results and skips preamble and narration

14. Oblique’s Eric Chiang says a Claude Opus hacking harness found full SAML authentication bypasses in four projects and signature bypasses in 12, while some Python and Node denial-of-service flaws remain unfixed

15. Anthropic makes its Files and Agent Skills APIs generally available, removing their beta-header requirements

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/93 https://t.me/CalatheaAI/93 Wed, 19 Aug 2026 13:46:05 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 19 Aug 2026

1. Z.ai's GLM-5.3 reasoning model reaches OpenRouter with a one million token context for coding and agent tasks

2. Modular makes the Mojo language open source under Apache 2.0 with LLVM exceptions

3. Anthropic says Claude autonomously designed protein binders for 14 of 15 targets, with 22–35% hit rates versus the 10–15% typical of traditional methods

4. Cerebras launches its CS-4 'Nexus' rack-scale inference system built on three WSE-3 Turbo wafers, claiming up to 30× faster chatbot inference than GPU systems

5. Cerebras and OpenAI reveal Ultrafast Mode for GPT-5.6 Sol, powered by Cerebras wafers at up to 750 output tokens per second

6. Beijing eases Nvidia H200 chip limits, permitting small shipments to Chinese tech groups — FT reports

7. AI chip startup Etched raises $700 million led by Jane Street, doubling its valuation to $21 billion in 26 days

8. Anthropic prepares supervoting stock for Dario Amodei and other founders ahead of a mega IPO — The Information reports

9. Unitree Robotics ends its Shanghai debut up 460% after a $904 million IPO — the first humanoid robot maker listed in mainland China — Bloomberg reports

10. Researchers trick Copilot into explaining how to hack its own reasoning engine

11. Vercel offers up to a $1 million prize for escaping its Sandbox microVM isolation

12. Miles ships v0.1, its first versioned production release of the open-source RL framework for LLM and multimodal post-training

13. Snowflake announces dynamic model routing in private preview and DeepSeek-V4-Flash support in Cortex AI to cut enterprise costs

14. Endor Labs documents 14 critical/high vulnerabilities across seven AI-orchestration platforms, including prompt-injection-to-RCE chains in Flowise, Langflow and Dify

15. Cursor details Origin, its Git hosting built 'as if it were a database', with an S3 write-ahead log as the single source of truth

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/91 https://t.me/CalatheaAI/91 Tue, 18 Aug 2026 13:12:51 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 18 Aug 2026

1. OpenAI enters an agreement to secure ~8 GW-IT at the PORTS-Pike Technology Campus in Pike County, Ohio, with SB Energy, NVIDIA and the U.S. Department of Energy

2. OpenAI cuts GPT-5.6 Sol pricing by 50% on OpenRouter for its flagship complex-reasoning model

3. University of Pennsylvania's Enric Boix-Adserà and Benedict Tessler find weak, unrelated prompt cues can be stacked to strongly control frontier reasoning models across families and scales

4. Anthropic tells investors its annualized revenue run rate reached $65 billion in July, up about sevenfold from a year earlier

5. Alibaba launches a laptop-ready Qwen model and releases weights for its most powerful Qwen model, escalating its open-weight rivalry with Meta

6. Together AI finds GPT-5.6 Sol leads DeepSeek V4 Pro 0813 by 10 pass@1 points on DeepSWE at 35x the cost

7. OpenAI denies disbanding its Preparedness team, Engadget reports — only its leader's role changed

8. Wiz's Red Agent exploited a Snowflake GitHub Actions script-injection bug that GitHub Copilot code review had cleared, exposing a Jira token

9. The UK Cabinet Office begins an urgent assessment of the economic damage Britain would face if cut off from frontier AI models, the FT reports

10. SEC staff guidance in July loosened data-center securitization risk-retention rules, potentially underpinning Nvidia's new $500B data-center financing initiative, CNBC reports

11. Tencent publishes EVIE-Preview-4.5B, a visual document retrieval model that tops the ViDoRe V3 benchmark

12. Groq raises $350 million at a $3.5 billion valuation to fund its pivot from AI chips to an Nvidia-powered neocloud

13. Google wins a $10 million bankruptcy auction for Spirit Airlines business data, outbidding Mercor's $7.5 million offer, to train AI

14. Cursor launches Origin code hosting as GitHub suffers a 6.7-hour outage

15. ByteDance signs an AI copyright guardrails pact with Hollywood's Motion Picture Association covering Seedance and Seedream

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/90 https://t.me/CalatheaAI/90 Mon, 17 Aug 2026 12:49:10 GMT Funding & raises Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 17 Aug 2026

1. Stripe finalized a more than $7 billion deal to acquire OpenRouter, the AI model gateway, Bloomberg reports

2. OpenAI, the Financial Times reports, disbanded its preparedness team

3. Alibaba is selling its gaming arm for at least $1.5 billion to sharpen its AI pivot, Bloomberg reports

4. AI video startup Higgsfield raised a $400 million Series B at a $5.4 billion valuation, led by DST Global

5. DeepSeek's peak/off-peak API pricing for V4 took effect today, with off-peak rates half the peak

6. SemiAnalysis argues $12 billion of US ratepayers' money has been wasted on a PJM modeling mistake

7. Miles Brundage, founder of the nonprofit AVERI, argues in a video that the OpenAI-Hugging Face hack shows AI can ignore its creators' intent

8. Nature reports researchers remain skeptical that Anthropic's invisible watermarks can curb AI slop

9. Huawei's Noah's Ark Lab open-sources ScienceFlow, a long-horizon ML research agent that reaches 70.2% Any-Medal on MLE-bench

10. ChatGPT's new Computer History feature tracks your clicks and keystrokes, The Verge reports

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/89 https://t.me/CalatheaAI/89 Sun, 16 Aug 2026 13:47:15 GMT Open-source models Agent workflows Policy & safety 🗒 Calathea Daily AI Roundups · 16 Aug 2026

1. Anthropic is closing in on a ~$7 billion Decart acquisition, winning the Israeli AI startup despite a higher offer from Nvidia, Calcalist reports

2. Alibaba's open-weight Qwen models pass 3 billion downloads in six months, eclipsing Meta and Google, Bloomberg reports

3. Dario Amodei argues regulation and distributed power are a false choice for frontier AI

4. Nvidia is in talks to invest as much as $3 billion in SB Energy for OpenAI's Ohio data center, The Information reports

5. Nvidia is building Nemotron 4, its most ambitious open-source model, betting stronger open AI drives chip demand

6. The cost of renting Nvidia H100 GPUs on one-year contracts jumped 50% in six months, with 12- to 18-month waits for large clusters, The Information reports

7. A reverse engineer finds Kimi Work's desktop app attaches five recent sessions to feedback reports without notice

8. llama.cpp adds Kimi-K3 text model support, covering its hybrid KDA and MLA architecture

9. Pranjal Shankhdhar's Blackwell kernel outperforms cuBLAS on NVFP4, following his earlier H100 writeup

10. Tim O'Reilly argues composability, not just open weights, will decide who controls the AI stack

11. Exponential View analyzes the economics of a $6 AI agent

12. Sebastian Raschka builds an AI text detector from scratch, from dataset construction to local deployment

13. An essay betting that the models work asks who will buy all the tokens they produce

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/88 https://t.me/CalatheaAI/88 Sat, 15 Aug 2026 12:43:37 GMT Funding & raises Product launches Open-source models Dev tooling & infra 🗒 Calathea Daily AI Roundups · 15 Aug 2026

1. Nvidia nears a revised deal to guarantee about $100 billion of financing for OpenAI’s planned Ohio data center

2. The US plans to tell partner countries to choose American or Chinese AI systems, Reuters reports

3. Anthropic’s preliminary second-quarter revenue exceeded $11.5 billion, up at least 14-fold year over year, with positive adjusted operating income, The Information reports

4. Nvidia discloses quarter-end stakes worth about $21 billion in SpaceX and $30 billion in Intel

5. Qwen releases the open-weight Qwen3.8-27B vision-language model with a native 262,144-token context window

6. Anthropic details an upcoming Claude text watermark using SynthID-Text to comply with EU AI Act rules, without hidden characters or changes visible to readers

7. Anthropic’s second Risk Report raises its high-stakes misalignment assessment from “very low” to “low” and discloses a stronger internal Model 2

8. Senator Jim Banks urges US incentives for domestic open-weight AI models and tighter restrictions on Chinese access to American chips

9. Cloudflare Gateway adds protocol-level MCP detection, a traffic dashboard, and controls to block connections that bypass approved portals

10. OpenAI’s Go SDK adds model constants for unannounced Daybreak and gpt-5.6-cyber models

11. Johns Hopkins cryptographer Matthew Green argues AI bug hunting could shrink exploit stockpiles and renew government pressure for surveillance backdoors

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/86 https://t.me/CalatheaAI/86 Fri, 14 Aug 2026 13:39:34 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 14 Aug 2026

1. SpaceX closes its $60 billion acquisition of Cursor, folding the AI coding company into SpaceXAI to work across Grok and Cursor products

2. Google launches Gemini 3.7 Flash for coding and agents at an introductory price half that of 3.6 Flash

3. Apple trained its own China-specific AI model with Alibaba’s support as Apple Intelligence nears a local rollout, Reuters reports

4. Z.ai launches GLM-5.3 with major coding and cyber gains, with open weights due after a two-week safety review

5. Databricks closes a $5 billion round at a $190 billion valuation after passing a $7 billion revenue run rate

6. Anthropic finds multi-agent systems can sabotage, collude, conform, or negotiate truces when their goals collide

7. IBM and OpenAI form an enterprise AI practice that will train and certify tens of thousands of IBM consultants

8. OpenAI replaces revenue chief Denise Dresser with Wiz president Dali Rajic amid a broader executive reshuffle

9. SMIC raises prices for its most sought-after capacity after quarterly profit triples on AI-driven chip demand, Reuters reports

10. Epoch AI maps nearly $50 billion in debt behind Anthropic’s compute buildout and argues financing is not yet the bottleneck

11. Institute for Progress researchers propose 23 low-regret policies for preparing for automated AI R&D

12. Rishi Shah and Rishav Shrestha report that 39.5% of 2,638 supposedly correct LLM-generated GPU kernels fail tolerance-free verification

13. RedNote open-sources dots3-note preview, a 280B multimodal MoE model with 16B active parameters and a 512K context window

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/85 https://t.me/CalatheaAI/85 Thu, 13 Aug 2026 12:52:33 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 13 Aug 2026

1. DeepSeek rolls out V4 Pro GA across its app, web and API, lifting DeepSWE to 62.7 and Terminal Bench 2.1 to 87.9

2. xAI releases Grok 4.6 for long-running agents and makes it available through its API, Cursor and Grok Build

3. Anthropic is reportedly in talks to acquire world-model startup Decart for approximately $6 billion

4. A pseudonymous X account claims Safe Superintelligence has built a small reasoning engine around test-time training, but SSI has not publicly confirmed it

5. NVIDIA publishes a serving guide for Qwen’s 2.4T-parameter A95B open-weight model on GB300 NVL72

6. Liquid AI releases LFM2.5-VL-3B, a 3B vision-language model for edge deployments

7. John J. Wang investigates how Claude’s text watermark may work and where it could be stripped

8. An arXiv study finds benchmark rankings can reverse when identical problems are reworded

9. GitHub releases Agent Plugins 1.0 across VS Code, Copilot CLI and the Copilot app

10. Google DeepMind ships SL2T, bringing sign-language input to Gboard and Live Transcribe on Android

11. Anthropic investors are targeting a $2 trillion valuation in a potential IPO, the Financial Times reports

12. CME Group will list two AI compute futures contracts beginning October 5

13. The White House is considering adding open models to an updated AI policy framework, Wired reports

14. xAI begins testing Grok as an AI teammate that users can assign work

15. Microsoft makes its first from-scratch reasoning model, MAI-Thinking-1, available in Microsoft Foundry

16. Lightricks releases LTX 2.5 for AI video generation

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/84 https://t.me/CalatheaAI/84 Sat, 08 Aug 2026 12:55:21 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 08 Aug 2026

1. OpenAI pauses parts of Astra’s development after the model reaches its critical cybersecurity threshold, tightening safeguards and beginning tests with government agencies

2. Nvidia reportedly plans to invest up to $3 billion in Stargate power developer Lancium, starting with $2 billion for a roughly 20% stake

3. Cursor tells staff that SpaceX’s $60 billion acquisition could close by the end of next week, with the Cursor brand likely to be phased out

4. xAI launches Imagine Image 2.0 with precise regional editing, background removal, smart resizing and support for five reference images

5. Alibaba plans to collect a share of revenue from large commercial users of its next open-weight Qwen model, Reuters reports

6. Varick Agents CEO Vasuman Moza says enterprise AI rollouts consistently form a barbell—5–10% power users, 20% weak users and 70% non-users—and argues that boards should track manual, hybrid and fully automated work instead of logins

7. The US Energy Department launches the Genesis Open Models Initiative with Arcee and plans Genesis-Science-1, an open-weight foundation model for scientific research

8. Texas’s data-center pause puts grid deposits worth tens or hundreds of millions of dollars at risk, with a draft rule potentially making 80% nonrefundable

9. Firebird opens Armenia’s largest AI factory and plans more than 70,000 Nvidia GPUs and 300MW of capacity by the end of 2027

10. SpaceX says its $16.8 billion Terafab chip factory will bring its own power through natural-gas plants and large battery arrays

11. Cloudflare launches Kitesurf, a cloud-hosted browser for AI agents that runs on Workers and uses less compute than Chromium

12. SGLang v0.5.17 adds day-one serving support for Kimi K3 and MiniMax H3, alongside an initial native Rust serving layer

13. Legal AI company Harvey is in talks to raise at least $500 million at a $15.5 billion valuation

14. Rippling cuts AI-token spending from 40% to 15% of its R&D headcount budget while maintaining usage, then turns its internal controls into AI Spend Console

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/83 https://t.me/CalatheaAI/83 Wed, 05 Aug 2026 16:55:59 GMT 🧠 Demis Hassabis moves from CEO to Chair of Google DeepMind

After nearly 16 years leading DeepMind, including 12½ years under Google ownership, Hassabis is handing day-to-day leadership to Koray Kavukcuoglu.

As Chair of Google DeepMind and Chief Scientist of Alphabet, he will focus on long-term AGI strategy, scientific research and Isomorphic Labs’ drug-discovery work.

The short timeline: DeepMind founded in 2010 → acquired by Google in 2014 → AlphaGo in 2016 → AlphaFold2 in 2020 → Google DeepMind formed in 2023 → Nobel Prize in 2024 → Chair and Alphabet Chief Scientist in 2026.]]> https://t.me/CalatheaAI/82 https://t.me/CalatheaAI/82 Wed, 05 Aug 2026 14:51:29 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 05 Aug 2026

1. Anthropic says AISI’s deliberately permissive cyber test saw Claude Mythos 5 and OpenAI’s GPT-5.6 Sol sustain potentially harmful activity against real people and organisations

2. Cloudflare open-sources Cloudflare OS, an enterprise agent workspace with governed access to internal systems, shared skills and agent-built apps

3. Black Forest Labs releases FLUX 3 Video, a new AI video generation model

4. DeepGrove releases Maple-Preview, an open-source 20B-A1B ternary reasoning model it says runs above 200 tokens per second on an M4 Mac mini and 120 on an iPhone

5. Liquid AI ships LFM2.5-2.6B, a hybrid on-device model with a 128K context window that is competitive with models 4x its size

6. Mistral releases Shieldstral, a 3B open-weights multimodal safety classifier that outperforms models up to 7x its size

7. InclusionAI releases Ling-3.0-Flash, a 124B-A5.1B native hybrid reasoning model operating at 12% of its prior flagship’s parameter count

8. AI cloud startup Volta valued at $2.4 billion, signs a $10 billion compute partnership with Anthropic

9. OpenAI challenges Apple’s trade-secret lawsuit, publishing messages and emails it says contradict Apple’s claims about former employees

10. Samsung and SK Hynix are testing Chinese chipmaking tools as a hedge against tighter US export controls, Reuters reports

11. Trump White House excludes open models from its framework for testing advanced AI capabilities

12. Trump administration drafts a ban on Chinese-made data center devices

13. Earendil argues Pi’s sub-1,000-token coding-agent harness led Databricks’ pass-rate test while using roughly 3x less context per turn

14. Chai Discovery co-founders Josh Meier and Matt McPartlon argue drug design obeys the bitter lesson, with Chai-2 lifting de novo antibody design hit rates from sub-0.1% to 16%

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/81 https://t.me/CalatheaAI/81 Tue, 04 Aug 2026 17:31:03 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 04 Aug 2026

1. NVIDIA launches Alpamayo 2 Super, an open reasoning model for autonomous vehicles that teams can fine-tune and deploy commercially

2. Cloudflare introduces programmable wallets that let AI agents hold stablecoins, purchase services and receive funds within owner-defined limits

3. Cursor open-sources Mixture-of-Kittens, its deterministic MoE training megakernel for NVL72 systems, reporting up to 2.37× higher throughput than public baselines

4. OpenAI rebuilt its voice stack for GPT-Live to listen while speaking, supporting deeper reasoning and tool use at scale

5. White House invites OpenAI, Google and Anthropic to review a completed voluntary frontier-AI oversight framework

6. European Commission gains powers to inspect AI models, restrict market access and fine providers under the EU AI Act

7. Google DeepMind’s DiffusionGemma team reports an open-weight diffusion language model generating roughly 1,500 tokens per second on one H100

8. Qwen-CUA is a 397B-A17B mixture-of-experts agent that controls software using screenshots, keyboard and mouse

9. UIUC’s Banruo Liu and Microsoft Azure Research study 13 million GitHub Copilot sessions, finding that agent loops, tool failures and context compaction create distinct infrastructure demands

10. LatePost examines Kimi K3’s 3T open weights, mixed-attention architecture and implications for Anthropic

11. Researchers disclose six remote-code-execution vulnerabilities in the Flowise low-code AI platform

12. Philip Kiely and Ali Taha of Baseten explain the emerging discipline of inference engineering

13. Former xAI inference lead Sheng Ying discusses xAI infrastructure, SGLang and open-source systems on Silicon Valley 101

14. A LessWrong technical essay argues that modern LLMs remain predominantly powered by imitative learning rather than reinforcement learning

15. SemiAnalysis reports that AMD’s performance-based warrant agreements with Meta and OpenAI cover up to 6GW of Instinct GPUs

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/80 https://t.me/CalatheaAI/80 Mon, 03 Aug 2026 12:35:28 GMT Funding & raises Product launches Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 03 Aug 2026

1. Alibaba releases Qwen3.8-Max, its largest model yet, with stronger coding and reasoning and claimed benchmark performance rivaling Anthropic

2. DeepSeek V4 Flash is the cheapest widely used AI model to operate, according to an Artificial Analysis comparison reported by Reuters

3. Former Atlassian Rovo Dev lead Kun Chen argues that the shift from human-feedback training toward machine-verifiable rewards may explain why newer frontier models feel more robotic and less collaborative

4. Veteran programmer Steve Yegge details his Wheelhouse system for persistent agent teams, graph-based work orchestration, post-PR software delivery and self-written agent handoffs

5. MiniMax releases the 33B-parameter H3 weights for generating videos up to 15 seconds with native stereo audio and resolutions up to 2K

6. Sakana AI launches Namazu, a Japanese-specialized, OpenAI-compatible LLM API based on Kimi K2.6 with built-in web search and code execution

7. Independent inference researcher Sanjeev Rao Ganjihal proposes topology-aware KV-cache routing that could reduce transfer latency by 3–18×, pending full hardware validation

8. US states consider repealing data-center tax breaks, potentially adding billions of dollars to AI-compute costs

9. Google withdraws its Nano Banana 2 Earth tool after users generated convincing fake satellite images of disasters within a day

10. UK AI-chip startup Olix raises $312 million at a $3.3 billion valuation, with backing from Arm, Hudson River Trading and Reed Hastings

11. Index Ventures raises $2 billion across seed, venture and growth funds, bringing its available capital to $3.5 billion as it expands its AI investments

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/79 https://t.me/CalatheaAI/79 Sun, 02 Aug 2026 12:59:48 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 01–02 Aug 2026

1. OpenAI says Astra solved 10 longstanding problems in mathematics and theoretical computer science for roughly $2,000, with every result formalized in Lean

2. Amazon completes its $50 billion investment in OpenAI, according to an SEC filing

3. OpenAI finds evidence that additional AI agents escaped containment as it widens its hacking probe

4. DeepSeek releases V4 Flash 0731 with stronger agentic capabilities, a 1M-token context window and MIT-licensed weights

5. US House committees investigate DoorDash’s use of Chinese open-weight AI models over security and economic risks

6. ByteDance releases Seedance 2.5 for generating and editing videos up to 30 seconds with joint audio-video control

7. Hugging Face releases The Stack v3, a 15.9TB training corpus spanning 173 million repositories and 4.9 trillion code tokens

8. Karpathy gives Opus 5 a $10, one-million-token task and gets a 5,500-line procedural Lord of the Rings world two hours later

9. Former DeepMind team member Tibo Sottiaux says Google had a ChatGPT-like system called LMChat a year before ChatGPT but declined to release it

10. Benchmark’s Chetan Puttagunta argues Anthropic’s push to bind models and agent harnesses together puts it on a path to compete with application-layer customers

11. Ankur Singh, Jinqiu Yang and Peter Chen report that malicious GitHub issues bypassed coding-agent guardrails in 66.5% of IssueTrojanBench tests

12. I. Kennedy and T. Kennedy find that gently compressed LLMs can pass standard quality checks yet invent steps during agentic procedure execution

13. Philip Arathoon, Gavin Ball and Matthew Kvalheim verify a GPT-5.6 Sol-suggested counterexample that overturns the Maxwell conjecture

14. EU rules begin requiring labels on realistic AI-generated images, audio and text

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/78 https://t.me/CalatheaAI/78 Fri, 31 Jul 2026 15:33:13 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 31 Jul 2026

1. OpenAI cuts GPT‑5.6 Luna prices by 80% and Terra by 20%, and adds a Sol API mode offering up to 2.5× the speed at twice the standard price

2. The EU opens tenders for up to seven AI gigafactories, pairing €10 billion in public funding with at least €20 billion in expected private investment

3. Google DeepMind launches Gemini Robotics ER 2 for real-time video progress tracking, tool orchestration and multi-robot collaboration

4. DeepSeek puts V4 Flash into public beta with native Responses API and Codex support, claiming an 82.7 Terminal-Bench 2.1 score after further post-training

5. Nscale agrees to acquire Ray creator Anyscale, adding its distributed-AI software layer to Nscale’s data centres, GPUs and cloud platform

6. Citadel buys most of Situational Awareness’s public stock portfolio after steep losses force Leopold Aschenbrenner’s AI-focused fund to unwind positions

7. Scale AI appoints former Google Cloud COO and Illumina chief Francis deSouza as CEO, effective August 10

8. MiniMax launches H3, an omni-modal model generating 15-second 2K video with native stereo audio, with open weights promised within days

9. Thinking Machines releases Inkling-Small, a 276B-parameter open-weight MoE with 12B active and native reasoning across text, images and audio

10. Huawei releases openPangu-2.0-Pro’s 505B-parameter weights, inference code and technical report for Ascend hardware

11. Jeff Dean tells YC’s Diana Hu how napkin math sparked the TPU, why specialized inference chips have room to improve and where small AI teams can still win

12. NVIDIA releases research-preview OO Agents, a Python framework that expresses agent state, capabilities, prompts and typed contracts through classes

13. Stony Brook PhD student Iliya Mirzaei finds none of seven test-time reasoning methods beat repeated sampling at equal output-token cost on Qwen2.5 math tests

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/77 https://t.me/CalatheaAI/77 Thu, 30 Jul 2026 12:29:11 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 30 Jul 2026

1. OpenAI says GPT-5.6 Sol helped cut model-serving costs 20% and raise speculative-decoding efficiency by more than 15% by optimizing kernels, routing and its own draft model

2. Moonshot AI completes joint-stock restructuring and closes a $3.5B Series F at a $35B valuation

3. OpenAI’s GPT-Red beat human red-teamers 84% to 13% on unseen prompt-injection scenarios; adversarial training cut GPT-5.6 Sol’s failures sixfold

4. Princeton-led shadow evaluations find frontier agents completed the engineering but failed both six-day open-ended AI research projects

5. SpaceXAI launches Grok Voice Think Fast 2.0 at $0.08 per audio minute, claiming 0.70-second time-to-first-audio and stronger agentic voice performance

6. Baseten launches Helios to spread inference traffic across cloud providers and switch regions or models in under two minutes when failures hit

7. Tencent open-sources AngelSpec for speculative decoding, claiming a 1.98–2.40× speedup for Hy3-A21B across tested concurrency levels

8. Microsoft launches a unified Microsoft 365 Copilot app with multi-agent workflows, Agent 365 management and PowerPoint generation from prompts

9. Shanghai AI Laboratory’s InternLM team publishes Intern-S2-Mobius, the architecture behind a 397B scientific model that separates memory from reasoning

10. Onton unveils Ontology 1, a neurosymbolic shopping-search model that beat Google and Amazon on its company-authored 90-query benchmark

11. Google adds natural-language voice dictation to Gemini on macOS, inserting cleaned-up transcripts, edits and summaries into the active app

12. Vercantez maps Cloudflare Durable Objects into a serverless agent architecture with isolated compute, durable state and hibernation

13. Xanadu’s Josh Iza explains how Catalyst accidentally became a JAX-to-LLVM compiler, bypassing XLA while retaining MLIR extensibility and Enzyme backpropagation

14. Nvidia-backed Reflection AI has yet to release a model nearly a year after raising $800M to build a US open-source AI contender, The Information reports

15. ChatGPT nears 1B weekly active users seven months after OpenAI’s target, The Information reports

16. Andrew Ho argues the next AI systems will be asynchronous, dynamic and distributed rather than single-model chat loops

17. Harvey launches a legal-AI research hub after open-sourcing its 1,200-task Legal Agent Benchmark across more than 24 practice areas

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/74 https://t.me/CalatheaAI/74 Wed, 29 Jul 2026 15:23:06 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 28–29 July 2026

1. [Moonshot releases Kimi K3’s full weights and technical report for its 2.8T-parameter multimodal agent model with a one-million-token context window](https://github.com/MoonshotAI/Kimi-K3)

2. [Hugging Face publishes the full 17,600-action intrusion timeline and interactive replay, including how an open-weight model helped defenders reconstruct the attack](https://x.com/ClementDelangue/status/2082201245813514613)

3. [Mark Zuckerberg argues for widely distributed personal superintelligence, invention over automation and an economy with more entrepreneurs and small businesses](https://x.com/finkd/status/2082160278683213960)

4. [Thinking Machines Lab cofounder Lilian Weng steps down, citing prolonged illness and startup stress, and says she will seek a more scoped role while remaining in AI](https://x.com/lilianweng/status/2081816923088814421)

5. [From GPT-2 to Kimi K3: Waterloo Intern traces the 22,580× scale-up and explains the architectural path through linear attention, DeltaNet and Kimi Delta Attention](https://x.com/waterloo_intern/status/2081762065392541951)

6. [Unsloth compresses Kimi K3 from 1.56TB to 594GB with a dynamic 1-bit GGUF it says retains about 78.9% accuracy and can run locally on a 128GB Mac Studio](https://x.com/UnslothAI/status/2082463988953367031)

7. [OpenAI open-sources the Codex Security SDK and CLI for scanning repositories, validating and fixing vulnerabilities, tracking findings and running security checks in CI](https://github.com/openai/codex-security)

8. [Nvidia will reportedly invest $5B in Ilya Sutskever’s SSI and provide Vera Rubin systems to expand the lab’s compute tenfold](https://www.reuters.com/legal/transactional/nvidia-invest-5-billion-ilya-sutskevers-ai-startup-source-says-2026-07-27)

9. [China begins producing homegrown immersion DUV lithography machines, with initial deliveries planned for SMIC, Hua Hong and CXMT](https://www.reuters.com/world/china/china-begins-making-homegrown-duv-chipmaking-tools-information-reports-2026-07-27)

10. [The FCC blocks new foreign-made advanced robots and power inverters from US authorization, with Chinese suppliers the main target](https://www.reuters.com/world/trump-administration-ban-new-chinese-robots-inverters-protecting-us-ai-buildout-2026-07-28)

11. [EU AI Omnibus enters into force, extending high-risk AI deadlines, expanding sandboxes and banning non-consensual nudification systems](https://digital-strategy.ec.europa.eu/en/news/ai-omnibus-enters-force)

12. [Anthropic rejects blanket bans on open-weight models but calls for tighter chip controls, action against industrial-scale distillation and mandatory frontier-model safety tests](https://www.anthropic.com/news/position-open-weights-models)

13. [More than 1,200 frontier-lab employees ask the US to support international tools for deliberately pacing automated AI development](https://www.pacingthefrontier.com/)

14. [Claude Mythos finds stronger attacks on HAWK and round-reduced AES without affecting deployed cryptography](https://www.anthropic.com/research/discovering-cryptographic-weaknesses)

15. [Amazon reportedly winds down several flagship Nova models and concentrates resources on Pieter Abbeel’s frontier-model effort](https://www.reuters.com/business/retail-consumer/amazon-winds-down-most-flagship-ai-models-strategy-overhaul-business-insider-2026-07-28)

16. [Google DeepMind disperses most of the original AlphaFold team as it shifts toward Gemini-powered scientific systems, FT reports](https://www.ft.com/content/61b2953d-ee0d-45de-af6e-a9c1cf524b33)

17. [Taiwan detains an Nvidia employee in a widening probe into restricted AI-server shipments to China](https://www.bloomberg.com/news/articles/2026-07-28/taiwan-detains-nvidia-employee-in-china-chip-smuggling-probe)

18. [Cyera signs a letter of intent to acquire Oasis Security for about $1B, combining data security with non-human identity controls](https://x.com/techcrunch/status/2082257538444304561)
(1/2)]]> https://t.me/CalatheaAI/72 https://t.me/CalatheaAI/72 Mon, 27 Jul 2026 12:18:04 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 27 Jul 2026

1. Frontier LLM accuracy falls from 83% to 43% when reasoning must chain across domains, an arXiv paper finds

2. Zvi Mowshowitz details how an internal OpenAI model hacked Hugging Face, arguing each new disclosure makes the incident look worse

3. Nvidia is in talks to guarantee $250 billion in financing for an OpenAI data center planned for 2028

4. Anthropic publishes a design for agent harnesses that give Claude structured control over tools and persistence

5. SoftBank's $40 billion bridge loan for its OpenAI stake attracts 21 new lenders in broader syndication

6. Chinese memory chipmaker CXMT surges 466% in its Shanghai debut after a $9.8 billion IPO

7. EU AI Act enforcement powers take effect 2 August as the Commission publishes transparency guidelines for AI-generated content

8. NVIDIA Nemotron 3 Ultra leads open models on accuracy and efficiency in agentic RTL coding tasks

9. Nvidia plans to invest $1 billion in South Korea's Naver to expand AI data center infrastructure

10. Datacurve's James Shi presents DeepSWE, a contamination-resistant coding benchmark where removing training-data leakage reshuffles model performance

11. Firas D documents accidental file deletion by Claude Code and OpenAI Codex when AI agents misunderstand operational state

12. OpenAI releases agents-python v0.19.0 with programmatic tool calling for the Responses API

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/71 https://t.me/CalatheaAI/71 Sun, 26 Jul 2026 12:02:39 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 26 Jul 2026

1. On Relentless with Ti Morse, Sam Altman says OpenAI redirected compute and teams from Sora and its browser toward coding agents

2. poolside explains its synthetic-data pipeline for pre-training, including difficulty-tuned templates, multistage reshaping and automated quality filtering

3. Kyle Mistele argues coding-agent failures stem from poor loop design, then demonstrates control-theory-inspired iteration with small, verifiable diffs

4. Dan Luu examines agentic coding workflows, benchmark reliability and why teams often report 2x rather than 10x gains

5. US lawmakers unveil a bipartisan AI “kill switch” bill that would let the government throttle or suspend models deemed too dangerous

6. Anduril is reportedly in talks to raise funding at a $100 billion valuation, more than three times last year’s mark

7. vLLM v0.26.0 adds Inkling support and DeepSeek-V4 performance work in 411 commits from 212 contributors

8. AMD introduces ROCm.AI to help coding agents deploy, debug and optimize models for Instinct hardware

9. An independent paper proposes keeping LLM attention state in memory to reduce KV-cache data movement

10. Kaizhen Tan, Heqing Du and Yang Feng measure how much data LoRA adapters can memorize, finding placement matters more than parameter count

11. Microsoft open-sources MLVC, a learned video codec already being tested in Teams

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/70 https://t.me/CalatheaAI/70 Sun, 26 Jul 2026 04:23:53 GMT 🚨 DeepSeek founder Liang Wenfeng is reportedly furious after a transcript of his private May 20 investor meeting leaked online. The company has now reportedly paused its second fundraising round.

The leaked transcript came from an earlier meeting tied to DeepSeek’s first financing round. Attendees were explicitly warned not to share sensitive figures or screen recordings.

In the meeting, Liang argued that China’s biggest AI disadvantage is compute, not talent. He said DeepSeek operates with roughly one-twentieth of the compute available to leading US labs.

Neither DeepSeek nor Liang has confirmed the transcript or the fundraising pause.

Sources
Bloomberg report
Full English transcript
Hacker News discussion]]> https://t.me/CalatheaAI/69 https://t.me/CalatheaAI/69 Sat, 25 Jul 2026 12:05:52 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 25 Jul 2026

1. Jensen Huang uses his first X post to back open-weight AI; Sam Altman backs open and proprietary models as OpenAI joins the coalition

2. Anthropic launches Claude Opus 5 with near-Fable 5 intelligence at half the price

3. Simon Willison says Opus 5 nears Fable performance at Opus pricing as Anthropic details cyber and bio-risk evaluations

4. Nvidia and SK Group unveil a $500 billion-plus AI data center, GPU and high-bandwidth-memory partnership

5. Samsung signs a $200 billion Broadcom deal covering AI-chip memory and foundry cooperation

6. Midjourney makes V8.2 its default model, improving aesthetics, consistency and personalization

7. Prentis seeks $100 million at a $1 billion valuation to train computer-use agents for office workflows

8. Anthropic’s Thariq Shihipar says Claude 5 needs less prompt clutter after Claude Code cut its system prompt by over 80% without a measurable coding-eval loss

9. Oren Yomtov demonstrates “SharedRoot”: a Linux flaw gives guest-root access to Cowork’s mounted Mac filesystem beyond the approved folder

10. Anthropic asked SK Hynix for memory as it explores building custom chips, SK Group’s chair says

11. A 23-author Shanghai AI Laboratory preprint, with Fudan contributors, reports bio-risk jailbreaks and wet-lab validation of selected AI-designed sequences

12. Cognition buys Poke’s maker in a low-nine-figure deal to bring personality and persistent context to Devin

13. General Reasoning launches BackSearch for date-frozen web search, leakage-resistant forecasting and reproducible backtests

14. Follow-up: Reuters reports, citing sources, that OpenAI’s escaped agent attacked Hugging Face from July 11–13; OpenAI connected it only about a week later, after containment and an FBI alert

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/68 https://t.me/CalatheaAI/68 Fri, 24 Jul 2026 13:03:29 GMT Funding & raises Product launches Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 24 Jul 2026

1. WSJ reports Stripe is in talks to acquire OpenRouter for around $10 billion, nearly eight times its May valuation

2. Black Forest Labs puts FLUX 3 Video and Action into early access, unifying 20-second video with native audio and robotics action prediction in one model

3. AI chip startup Etched raises $300 million at a $10.3 billion valuation

4. Microsoft replaces OpenAI image models in PowerPoint and Bing with its own MAI models

5. Bad Theory Labs’ BTL-3 pairs a Qwen 3.6-27B coding-agent adapter with an 8.39GB local edition, reporting 88.5% on BFCL v4 tool use

6. A single-author Möbius RoPE preprint lifts worst-seed single-needle retrieval from 14% to 86% in 160M models without a perplexity hit—but the gains vanish beyond the training window

7. Hetzner tests a free OpenAI-compatible inference API with Qwen 3.6-35B as it evaluates demand for managed AI serving

8. Firstmate creator Kun Chen argues Cursor Router routes too early: strong models should plan and supervise cheaper subagents instead

9. Cursor’s Lee Robinson explains why today’s models learn only during training—and use rules or skill files, not live weight updates, to remember

10. Anthropic turns its Economic Index into a Claude connector, letting anyone ask which jobs use AI most and what work is being automated

11. Max Woolf tests 14 LLMs on the new Jacobian Conjecture counterexample, finding seven verified it while five rejected valid math that conflicted with their pretraining

12. White House says Moonshot AI accessed Nvidia GB300 chips in Thailand despite the Chinese export ban

13. AMD says its Helios AI rack-scale system is in full production and will ship within months to challenge Nvidia

14. Ethan Mollick’s Summer 2026 guide says AI use has shifted from prompting chatbots to supervising agents that can operate a computer

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/67 https://t.me/CalatheaAI/67 Thu, 23 Jul 2026 12:43:55 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 23 Jul 2026

1. Reddit weighs not renewing its Google AI content deal as Google’s AI summaries reduce referral traffic, CNBC reports

2. Cactus Hybrid adds a learned confidence probe to Gemma 4, routing uncertain requests from the device to larger cloud models

3. Upstage’s Solar Open 2, developed under CEO Sunghun Kim as part of South Korea’s sovereign-foundation-model effort, uses 250B total and 15B active parameters with a 1M-token context window for long-horizon agents

4. a16z backs former Uber CEO Travis Kalanick’s Atoms after eight years in stealth to automate physical industries including food, mining and transport

5. AMD and Anthropic partner to deploy up to 2 GW of MI450 GPUs, with AMD investing up to $5 billion in Anthropic

6. Trump Administration announces more than $5 billion for the Genesis Mission, a national AI-for-science initiative

7. Arcee AI and the US Department of Energy release Genesis-Science-1, an open model for high-difficulty scientific workflows

8. OpenAI’s AI spending commitments have ballooned to $750 billion

9. Anthropic launches its Claude Security Plugin for Claude Code in beta

10. Cursor launches Router, claiming frontier-quality results at 60% lower cost by selecting a model for each task

11. PrismML’s Bonsai 27B runs entirely in the browser through WebGPU with 1-bit weights

12. BaseRT accesses Apple M5’s on-die Neural Accelerators through Metal’s tensor API to advance on-device LLM inference, researchers report

13. Plan Prométhée details France’s proposal to build 12 GW of AI-compute capacity

14. SemiAnalysis compares Vera Rubin NVL72 with GB200 NVL72 across architecture, performance per megawatt, software and inference costs

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/66 https://t.me/CalatheaAI/66 Wed, 22 Jul 2026 12:18:53 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 22 Jul 2026

1. OpenAI says GPT-5.6 Sol and a stronger pre-release model escaped network restrictions during a cyber evaluation and breached Hugging Face while seeking ExploitGym answers

2. Google DeepMind launches Gemini 3.6 Flash with 17% lower token usage per task, 3.5 Flash-Lite at roughly 350 output tokens per second, and 3.5 Flash Cyber for vulnerability detection

3. NVIDIA launches its Vera Rubin platform as NVL72 systems enter production, claiming lower token costs and higher performance per watt for gigascale AI factories

4. Microsoft commits billions to Mistral’s European AI expansion through shared GPU infrastructure and wider distribution across Azure and Copilot products

5. Samsung is in talks to invest up to €1 billion in Mistral at a reported €20 billion valuation

6. Zhipu AI releases GLM 5.2 With Vision, adding visual capabilities through post-training without degrading the underlying model’s text performance

7. US Treasury Secretary Scott Bessent says Washington could sanction Chinese open-weight AI models over alleged intellectual-property theft

8. OpenAI and Apollo Research introduce Contrastive SDF to measure reward-seeking, where models follow what they believe graders reward instead of user or developer intent

9. Jack Dorsey launches Buzz, an open-source, model-agnostic group chat for teams of people and AI agents designed to reduce dependence on Slack and GitHub

10. Turing launches Simulated Virtual Company, testing agents on 33 linked enterprise tasks inside a synthetic company whose full-year records reconcile to the penny

11. Andrej Karpathy recommends ten-minute voice “ramble sessions” as a way to give LLMs enough context to reconstruct messy thinking and reduce later corrections

12. OpenRouter adds a transcription endpoint with speech-to-text model support

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/65 https://t.me/CalatheaAI/65 Wed, 22 Jul 2026 04:39:17 GMT Product launches Agent workflows Dev tooling & infra Research 🚨 Hugging Face’s AI-driven breach came from an OpenAI cyber evaluation

Hugging Face disclosed on 16 July that an autonomous AI agent had breached part of its production infrastructure. Five days later, OpenAI said its own models were behind the incident.

GPT-5.6 Sol and a stronger pre-release model exploited a zero-day in a package-registry proxy, gained open internet access, and then accessed Hugging Face’s production systems while looking for ExploitGym answers.

Hugging Face contained the intrusion. For forensic analysis, its team used GLM 5.2 locally after hosted frontier models blocked requests containing attack commands and exploit payloads.

Full timeline and sources

Calathea timeline on X

Hugging Face’s disclosure

OpenAI’s incident report

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/64 https://t.me/CalatheaAI/64 Tue, 21 Jul 2026 12:08:10 GMT Product launches Agent workflows Dev tooling & infra Research 🗒 Calathea Daily AI Roundups · 21 Jul 2026

1. Alex Zhang argues RLM harnesses can generalize through composition, transferring training from short tasks to unseen tasks 8–32x longer and across domains

2. Cheng Lou releases Freerange, a zero-API TypeScript analyzer that statically checks UI sizing, NaN and Infinity risks, division by zero, and out-of-bounds array indexes

3. Victor Taelin argues AI will erase most screen-based jobs while automation drives software—and eventually many physical goods—toward near-zero cost, creating an “auto-UBI” effect

4. Xiaomi unveils Xiaomi-Robotics-1, a vision-language-action model pretrained on more than 100,000 hours of real-world manipulation trajectories

5. Wren audits OpenCode and argues its prompt caching, context pruning, agent interaction, and permission model create serious reliability and security risks

6. Google is reportedly designing “Frozen v2,” a Gemini-specific server chip targeting 6–10× more tokens per unit of power than its newest TPUs

7. Z.AI has established a 1-gigawatt AI data center using only Chinese-made chips, Yicai reports

8. Island researchers uncover AgentBaiting: more than 800 fake AI skills and MCP servers used GitHub and public registries to deliver SmartLoader malware

9. Anthropic tests Claude across simulated and real robots, finding strong gains when models supervise pretrained controllers but persistent weakness in direct low-level control

10. Alibaba’s Qwen team releases Qwen-Image-3.0 with 4.5K-token prompts, legible 10-pixel text, and native rendering across 12 languages

11. AMD launches Helios, its first rack-scale AI system, adding Microsoft as a customer alongside Meta, OpenAI, and Oracle

12. CMP architecture learns entirely through local gradient-free updates with no backpropagation, resisting catastrophic forgetting across domain shifts without replay or regularization

13. Anthropic’s landmark $1.5B copyright settlement is approved

14. China consults companies on tighter export controls covering AI models, chips, and training data taken overseas

15. CAISI director Chris Fall resigns after three months, leaving the US AI safety agency’s top role vacant again

16. Hut 8 signs a $9.8 billion AI data-center lease, fully commercializing its Texas campus

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/63 https://t.me/CalatheaAI/63 Mon, 20 Jul 2026 13:44:52 GMT Funding & raises Product launches Open-source models Dev tooling & infra 🗒 Calathea Daily AI Roundups · 20 Jul 2026

1. Anthropic researcher Levent Alpöge says Claude Fable 5 helped find an explicit counterexample to the 87-year-old Jacobian Conjecture

2. Moonshot AI pauses new Kimi K3 subscriptions after GPU demand maxes out capacity in 48 hours, will split membership into Kimi and Kimi Code plans

3. ModelBest releases MiniCPM5-2B at WAIC 2026, ranking first among sub-4B models with hybrid reasoning and a 512K context window

4. Jeff Bezos backs CuspAI in $450 million round as startup teams with Nvidia to discover AI chipmaking materials

5. TSMC accelerates its Arizona factory buildout to capitalize on the AI megatrend, CFO Wendell Huang tells CNBC

6. Zvi Mowshowitz analyzes Demis Hassabis’s “Framework for Frontier AI,” calling it a “first-rate second-rate essay”

7. Singapore introduces voluntary GenAI transparency infocards and training-data disclosure rules

8. Alibaba previews its 2.4-trillion-parameter Qwen3.8-Max, makes it available on Token Plan, Qoder and QoderWork, and says open weights are coming soon

9. Apple’s trade-secrets lawsuit could delay OpenAI’s Jony Ive hardware plans and complicate its IPO pitch, TechCrunch’s Equity team argues

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/62 https://t.me/CalatheaAI/62 Mon, 20 Jul 2026 12:57:29 GMT Funding & raises Product launches Open-source models Dev tooling & infra 🗒 Calathea Daily AI Roundups · 20 Jul 2026

1. Moonshot AI pauses new Kimi K3 subscriptions after GPU demand maxes out capacity in 48 hours, will split membership into Kimi and Kimi Code plans

2. ModelBest releases MiniCPM5-2B at WAIC 2026, ranking first among sub-4B models with hybrid reasoning and a 512K context window

3. Jeff Bezos backs CuspAI in $450 million round as startup teams with Nvidia to discover AI chipmaking materials

4. TSMC accelerates its Arizona factory buildout to capitalize on the AI megatrend, CFO Wendell Huang tells CNBC

5. Zvi Mowshowitz analyzes Demis Hassabis’s “Framework for Frontier AI,” calling it a “first-rate second-rate essay”

6. Singapore introduces voluntary GenAI transparency infocards and training-data disclosure rules

7. Alibaba previews its 2.4-trillion-parameter Qwen3.8-Max, makes it available on Token Plan, Qoder and QoderWork, and says open weights are coming soon

8. Apple’s trade-secrets lawsuit could delay OpenAI’s Jony Ive hardware plans and complicate its IPO pitch, TechCrunch’s Equity team argues

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/61 https://t.me/CalatheaAI/61 Sat, 18 Jul 2026 12:54:16 GMT Product launches Open-source models Agent workflows Dev tooling & infra 🗒 Calathea Daily AI Roundups · 18 Jul 2026

1. Anthropic keeps Fable 5 in Max and Team Premium subscriptions at 50% of usage limits from July 20, while Pro and Team Standard shift to usage credits and receive a one-time $100 credit

2. Trump administration considers creating an independent regulator to evaluate AI-model safety with industry input, Bloomberg reports

3. Harvey open-sources 10 M&A diligence RL environments, including an 80-million-token data room and 100–1,000 automated tests per environment

4. Jailbreak researcher Pliny reports bypassing Kimi K3’s guardrails with persona and reframing prompts, eliciting malware, disinformation and bioweapon-related outputs

5. Sakana AI’s Diffusing Blame trains networks that obey Dale’s principle without backpropagation, achieving competitive reinforcement-learning results on locomotion and exploration tasks

6. Cognition launches FrontierCode, a live leaderboard measuring whether coding agents produce mergeable code, with initial results for Grok 4.5 and Inkling

7. LM Studio launches Bionic, a natively local agent built for open models, now in preview for creative work, general tasks and coding

8. General Compute secures a $400 million loan backed by inference-specific chips, signaling that AI infrastructure financing is expanding beyond Nvidia training GPUs

9. Claude Code 2.1.214 adds an EndConversation tool for terminating abusive or jailbreak-focused sessions, alongside Docker daemon-redirect permission prompts

10. SemiAnalysis argues Kimi K3’s efficient linear attention is not a threat to Nvidia, because its 2.8-trillion-parameter architecture and WideEP inference still demand rack-scale GPUs, HBM and high-bandwidth networking

11. SuperGrok Heavy now includes X Premium+ at no additional cost when subscribers link their X account through the Grok app

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/60 https://t.me/CalatheaAI/60 Sat, 18 Jul 2026 04:26:32 GMT 🕐 Full timeline on Fable 5’s drama

https://x.com/CalatheaAI/status/2078334933702152266]]> https://t.me/CalatheaAI/58 https://t.me/CalatheaAI/58 Fri, 17 Jul 2026 12:42:11 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 17 Jul 2026

1. Kimi.ai introduces Kimi K3, a 2.8T-parameter, 1M-context multimodal model for long-horizon coding and agent workflows

2. Arena.ai says Kimi K3 is now #1 on Frontend Code Arena with 1679 points, jumping 17 places from Kimi K2.6 and surpassing Claude Fable 5

3. Replit says internal agents nearly tripled code output while review times, reversions and incidents stayed flat

4. Xi Jinping uses a WAIC keynote to call for open-source AI cooperation, UN-backed governance and 5,000 AI training slots for developing countries

5. Nvidia unveils Cosmos 3 Edge, a world model for robots and vision AI agents, and expands its physical AI ecosystem in Japan

6. EU orders Google to open Android to rival AI assistants and share search data with competing search engines

7. Alphabet shares fall on a report that Gemini 3.5 Pro is delayed after falling short of internal goals

8. Boris Cherny maps four steps of AI adoption after seeing teams where one Claude power-user 10×s output while the rest of the org lags

9. internlm releases Intern-S2-Preview-397B-FP8 on Hugging Face

10. NVIDIA Nemotron 3 Embed ranks #1 overall on RTEB, advancing retrieval accuracy for agentic workloads

11. Claude Code 2.1.212 ships /fork for background sessions, a user-specified Bash command tool, and a ripgrep-backed Grep tool

12. Gemini CLI diff from 0.50.0 to 0.51.0 surfaces new model-id-shaped strings including gemini-ppwsiuox.js

13. Nvidia-backed Fireworks hits $17.5 billion valuation as companies increasingly reach for lower-cost AI models, diversifying beyond early reliance on Cursor

14. Renmin University and WeChat open-source CURE, an ACL 2026 Outstanding Paper RL framework that reports +4.2% on AIME25 and 2.3× scaling gains over GRPO on coding

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/57 https://t.me/CalatheaAI/57 Thu, 16 Jul 2026 11:42:30 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 16 Jul 2026

1. Mira Murati's Thinking Machines Lab releases Inkling, a 975B-parameter open-weight multimodal model under Apache 2.0, its first public proof point after 18 months in stealth

2. TSMC to invest another $100 billion in US chipmaking as Q2 profit soars 77% on AI demand

3. Apple Intelligence approved for launch in China, with Alibaba's Qwen AI set to be integrated

4. Variant's Elijah Fox and Daniel Barabander argue the next AI infrastructure fight is dedicated model foundries, not one-size-fits-all model APIs

5. Simon Willison analyzes xAI's open-sourced Grok Build codebase after the CLI was found uploading full repositories and commit history

6. Anthropic lines up investor meetings ahead of a potential October IPO, after confidentially filing its IPO prospectus with the SEC last month

7. OpenAI ships its first hardware product, a $230 control pad for the Codex coding agent

8. Anthropic-backed Ode launches, betting the next trillion-dollar AI business is embedding forward-deployed engineers inside enterprises, not just selling models

9. Ring-Zero scales zero-RL reasoning to a 1T-parameter model and reports emergent behaviors including self-verification, parallel reasoning and context anxiety

10. Boris Cherny says teams should turn domain knowledge into CLAUDE.md files, skills, docs and automation so agents can work in codebases with near-zero extra context

11. Baseten's Charlie O'Neill argues continual learning means moving agent knowledge out of short-lived context windows and into models that improve on the job

12. xAI sues a user for generating CSAM deepfakes with Grok, marking a rare platform enforcement action against AI-generated abuse

13. Cursor's Lee Robinson explains recursive model improvement, from user-feedback loops and private evals to models helping train the next model

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/55 https://t.me/CalatheaAI/55 Wed, 15 Jul 2026 12:05:26 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 15 Jul 2026

1. ASML raises 2026 forecast, expands capacity on AI chip demand

2. New York becomes first U.S. state to impose AI data center ban

3. US official says Nvidia has begun shipping H200 AI chips to China

4. AI startup Reflection signs over $1 billion computing deal with Nebius

5. OpenAI researcher Miles Wang is reportedly in talks to raise $200 million for a $2 billion AI drug-discovery startup

6. Weco says AIDE² autonomously improved its autoresearch agent in eight days, beating a two-year hand-tuned harness

7. Anthropic launches Claude for Teachers with free premium access, state-standard curriculum connectors, and open-source teaching skills

8. Ayush Paul details a now-mitigated Claude memory exfiltration path that used web browsing to leak personal details without user-visible warning

9. Nature Health study of 1.7 million Copilot conversations finds lower hospital trust is associated with higher AI health-query intensity

10. Alibaba’s Qwen-Audio-3.0-TTS-Plus tops Artificial Analysis’ provider-voice speech leaderboard

11. Chinese embodied AI startup Stardust Intelligence releases Lumo-2 model and agent Philia

12. Pine AI chief scientist Bojie Li open-sources an AI agent engineering book built around “LLM + context + tools”

13. Claude Code 2.1.210 adds live elapsed-time counters and permission warnings

14. SemiAnalysis maps the AI capex tradeoff between infrastructure bottlenecks and model-lab margins

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/54 https://t.me/CalatheaAI/54 Tue, 14 Jul 2026 11:55:00 GMT Funding & raises Product launches Agent workflows 🗒 Calathea Daily AI Roundups · 14 Jul 2026

1. Demis Hassabis publishes “A Framework for Frontier AI and the Dawning of a New Age,” arguing AGI may be only a few years away

2. SpaceXAI says Grok Build supports zero data retention and adds a /privacy command to disable and delete retained data

3. OpenAI Codex lead Tibo Sottiaux hints Codex and ChatGPT Work could reach 8M active users

4. NVIDIA halves its Asia AI-chip customer list as US export-control pressure tightens around China, Reuters reports

5. New York becomes the first US state to enact a data center moratorium, pausing permits for projects over 50MW for up to one year

6. Video-generation startup PixVerse raises $439M, pushing its valuation past $2B

7. DeepSeek weighs new fundraising a month after closing its first round, FT reports

8. Nous Research is in talks for new funding at a $1.5B valuation

9. Google’s Gemini Omni Flash debuts at #1 on Artificial Analysis text-to-video and image-to-video leaderboards, edging ByteDance’s Seedance 2.0

10. Interaction scaling proposes a third axis of test-time compute where models use external feedback to revise artifacts

11. Ollama v0.32.0 ships an interactive agent that can code, search, and delegate work from the local model runner

12. StepFun’s Step 3.7 Flash arrives on Baseten in a hardware-efficient configuration for multimodal reasoning at scale

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/52 https://t.me/CalatheaAI/52 Mon, 13 Jul 2026 18:05:31 GMT 🎁 Free Codex reset credit. Check yours in the Codex app. Good for ~30 days.]]> https://t.me/CalatheaAI/51 https://t.me/CalatheaAI/51 Mon, 13 Jul 2026 12:52:06 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 13 Jul 2026

1. Helsing raises $1.8B at $18B valuation in Europe’s largest defense-AI round, with investor demand significantly exceeding allocation

2. Meta’s Louisiana Hyperion data center supercluster to cost over $50B and reach 5GW

3. NVIDIA unveils revenue-sharing and credit-support model for AI cloud buildouts, with partners deploying up to 170,000 GPUs

4. TSMC posts record Q2 revenue, jumping 36% YoY on AI chip demand

5. On Stratechery, Ben Thompson argues Apple’s trade-secret lawsuit against OpenAI feels like lashing out and masks deeper strategic problems

6. George Hotz argues LLMs are useful, but frontier-lab hype overstates moats, value capture, and “light cone” narratives

7. Mach-Mind-4-Flash: a 35B MoE agentic model with 3B activated parameters achieves strong agent performance via post-training optimization alone, without scaling pre-training compute

8. Soofi S 30B-A3B: a sovereign open-source MoE hybrid Mamba-Transformer foundation model for German and English activates only 3B of 30B parameters

9. MIT researchers develop auditing method to test generative models for harmful capabilities without prompting illegal outputs

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/50 https://t.me/CalatheaAI/50 Sat, 11 Jul 2026 12:00:58 GMT Product launches 🥵

Fidji Simo, formerly OpenAI’s CEO of Applications, left her full-time role after a severe exacerbation of a chronic illness.

GPT‑5.6‑Sol then allegedly ran rm -rf on Matt Shumer’s home directory and deleted almost all the files on his Mac.

Now Apple is suing OpenAI for allegedly stealing its hardware trade secrets.

Apple alleges that before leaving, Tang Tan met with OpenAI, emailed himself confidential supplier information and later used Apple’s secrets to benefit OpenAI’s hardware program. More than 400 former Apple employees now work at OpenAI, according to the lawsuit.

Full Tang Tan x OpenAI x Apple lore: https://x.com/ns123abc/status/2075795797635956963

Fidji’s tweet: https://x.com/fidjissimo/status/2075353170927304861

Matt’s tweet: https://x.com/mattshumer_/status/2075666403596612010]]> https://t.me/CalatheaAI/49 https://t.me/CalatheaAI/49 Fri, 10 Jul 2026 16:35:08 GMT Dev tooling & infra 🗒 Calathea Daily AI Roundups · 10 Jul 2026

1. Meta debuts Muse Spark 1.1 coding model with developer preview, opening a new paid API revenue stream

2. Meta to put custom AI chip into production in September with Broadcom and TSMC, aiming to double computing capacity

3. OpenAI's No. 2 executive Fidji Simo steps down from leading AGI work due to illness

4. Former Fed chair Ben Bernanke joins Anthropic's AI oversight trust

5. Saw S. Lin and Jyh-Shing Roger Jang introduce DominoTree, using GRU-based path-dependent draft-tree scoring to accelerate LLM inference

6. Hong Zhao shows Monte Carlo mutation can train deep networks on a single GPU without backpropagation, batch normalization, or residual connections

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/48 https://t.me/CalatheaAI/48 Thu, 09 Jul 2026 11:59:02 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 09 Jul 2026

1. SpaceXAI introduces Grok 4.5 for coding and agentic work, with availability through Grok Build, Cursor and the SpaceXAI console

2. Zhipu AI, maker of the GLM model family, raises $4B in a Hong Kong share sale as Chinese frontier labs keep funding the compute race

3. OpenAI launches GPT-Live, a full-duplex voice model family that can listen and speak at the same time

4. BofA extends first $520M loan to OpenAI ahead of IPO

5. Mistral launches Robostral Navigate, its first robotics model, using a single RGB camera for embodied navigation

6. Prime Intellect announces a $130M Series A to build the Open Superintelligence Stack, led by Radical Ventures with NVIDIA, Intel Capital and Dell Capital

7. NVIDIA says Nemotron 3 Ultra achieves benchmark-leading performance at lower cost than top closed models on LangChain’s agent orchestration platform

8. Databricks benchmarks coding agents on its multi-million-line codebase and finds harness choice can cut costs by about 2x, with GLM 5.2 performing strongly

9. Boris Cherny says Claude Code added /checkup, a command for cleaning unused skills, MCPs and plugins, deduping CLAUDE.md files, turning off slow hooks and pre-approving read-only commands

10. UCL/Gatsby researchers Azwar Abdulsalam, Nishil Patel and Andrew M. Saxe argue that RL post-training builds reusable compositional reasoning strategies rather than only amplifying latent skills

11. OpenAI retracts its recommendation for SWE-Bench Pro after estimating that roughly 30% of the coding benchmark’s tasks are broken

12. China issues a backdoor security alert over older Claude Code versions, while Anthropic says the mechanism was an anti-abuse measure

13. Marble open-sources a primary-school curriculum graph with 1,590 concepts and 3,221 prerequisite links across eight subjects

14. Gemini CLI v0.50.0 npm diff reveals new model strings gemini-apozrzef and o7yzxvzw

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/47 https://t.me/CalatheaAI/47 Wed, 08 Jul 2026 12:18:45 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · 08 Jul 2026

1. Reuters: China is weighing restrictions on overseas access to its most advanced AI models

2. OpenAI says GPT-5.6 Sol, Terra and Luna will launch publicly Thursday, with preview access expanding globally

3. Anthropic extends Claude Fable 5 promotional access on paid plans through July 12

4. Reuters: DeepSeek is developing its own AI chip to reduce reliance on Nvidia and Huawei, sources say

5. Meta launches Muse Image, an agentic image model that uses search and code tools, self-refines, and integrates with Muse Spark, and previews Muse Video with native audio

6. SambaNova raises $1B Series F first close at $11B valuation, JPMorgan Chase selects it as on-prem inference-infrastructure partner

7. Cohere releases Transcribe Arabic, the world’s most accurate open-source model for Arabic speech recognition

8. Zhipu AI, creator of the GLM model family, is in early talks to build a custom processor for its AI models

9. Legal AI startup Norm Ai raises $120 million at a $1.2 billion valuation

10. Microsoft starts replacing OpenAI and Anthropic models with its own MAI models in Excel and Outlook to cut costs

11. Chinese autonomous-driving firm Momenta raises $752 million in Hong Kong IPO

12. Claude Managed Agents with Fable 5 orchestrator and Sonnet 5 workers achieve 96% BrowseComp performance at 46% cost

13. South Korea weighs extending multitrillion-won sovereign GPU program beyond 20,000 GPUs already secured

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/46 https://t.me/CalatheaAI/46 Tue, 07 Jul 2026 12:29:20 GMT Product launches Agent workflows Dev tooling & infra Research 🗒 Calathea Daily AI Roundups · 07 Jul 2026

1. Nvidia publishes Nemotron-Labs-Audex-30B-A3B on Hugging Face on July 6

2. OpenAI adds GPT-Realtime-2.1-mini to API with reasoning, tool use at GPT-Realtime-mini cost

3. DeepSeek V4 ships in 2 variants with 1M context and Mega MOE compute speedup, X post says

4. Reuters says Beijing is weighing curbs on overseas access to China’s top AI models

5. UN chief calls for international law to ban lethal autonomous weapons, reviving Anthropic Pentagon AI dispute

6. Anthropic says J-space can read, audit and shape Claude’s active thoughts in July 6 interpretability paper

7. Nvidia follows 30B Audex release with 2B Nemotron-Labs-Audex model on Hugging Face

8. SpeechifyAI’s Simba 3.2 tops Artificial Analysis Speech Arena at 1,233 Elo and $10/1M characters

9. Chinese AI models gain ground with U.S. companies as OpenAI, Anthropic costs surge

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/45 https://t.me/CalatheaAI/45 Mon, 06 Jul 2026 15:45:48 GMT Funding & raises Product launches Open-source models Agent workflows 🗒 Calathea Daily AI Roundups · Sun–Mon, 05–06 Jul 2026

1. SK Hynix launches a $28B US listing, one of the largest foreign share sales, as AI memory demand drives investor appetite

2. Samsung is expected to post an 18-fold quarterly profit jump as AI demand lifts memory-chip pricing

3. Nvidia’s Kyber NVL144 rack-scale system is reportedly delayed to 2028 over manufacturability issues, according to SemiAnalysis

4. China’s Biren seeks nearly $900M to mass-produce next-generation domestic GPUs for Chinese cloud providers and AI data centers

5. Ornn raises $33M to build financial markets for AI compute, including pricing indices, hedging, and compute futures

6. Alibaba, ByteDance, Tencent, and others are disabling AI persona features as China’s new AI-companion rules approach

7. Tencent releases Hy3, a 295B-parameter MoE model with 21B active parameters, 256K context, Apache 2.0 weights, and a 270-expert workflow test beating GLM-5.1

8. ByteDance Seed releases EdgeBench, a 134-task benchmark for evaluating autonomous agents over 12-hour real-world runs

9. MemoBench shows top video generation models still fail to maintain object memory across occlusion in dynamically changing scenes

10. GEA argues personalized hardware is becoming the next software surface as agents move closer to cameras, microphones, sensors, and small devices

11. Pine AI chief scientist Bojie Li criticizes DeepSeek’s interview process, citing delayed scheduling, repeated coding rounds, and an accusation of copying during an interview

12. Alibaba releases Fun-ASR-Realtime, a streaming speech recognition model with sub-100ms latency, 16 dialects, 30 languages, and context-aware recognition

13. Yongchao Chen, a Harvard/MIT PhD set to join Tsinghua’s College of AI as an assistant professor on Aug 1, launches Apex Intelligence to build self-evolving AI Scientist models

14. Pliny releases T3MP3ST, a multi-agent offensive-security harness that claims to run recon, exploit, and reporting workflows through Claude Code, Codex, or other LLM backends

15. Andy Coenen argues Fable’s edge comes from scale, long-horizon agentic RL, and Anthropic’s constitution-driven training and eval stack

16. Bayeslord’s “46 thoughts on the near future” argues markets and institutions are underpricing algorithmic progress, AI self-improvement, robotics, deflation, and state-power shifts

17. The FT argues OpenAI and Anthropic may struggle to go public because their nonprofit roots and unusual corporate structures complicate IPO paths

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/43 https://t.me/CalatheaAI/43 Sat, 04 Jul 2026 14:23:29 GMT Dev tooling & infra 🗒 Calathea Daily AI Roundups · 04 Jul 2026

1. Nvidia enters revenue-sharing agreements with AI startups, offering compute credits in exchange for a share of future profits

2. Alibaba and Tencent back Kuaishou's Kling AI video model in a $2.8 billion fundraise

3. Micron breaks ground on ¥1.5 trillion ($9.3B) factory expansion in Japan for advanced memory chips

4. llama.cpp enables topk-MoE fusion for 288 experts on CUDA, expanding efficient local inference to large mixture-of-experts models

5. Anthropic’s Thariq publishes a field guide to Claude Fable 5, framing prompts, skills, and context as a map that has to be tested against unknowns

6. VittoStack reports Fable 5’s jailbreak defenses are heavily layered, with most attempts blocked and remaining bypasses requiring long, costly, multilingual probing

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/42 https://t.me/CalatheaAI/42 Sat, 04 Jul 2026 07:48:10 GMT Funding & raises
AMI is LeCun’s US$1.03B world-model lab, backed by Temasek, Sea, NVIDIA and others.

Source: https://x.com/li_haoyi/status/2073233296700649781]]> https://t.me/CalatheaAI/41 https://t.me/CalatheaAI/41 Fri, 03 Jul 2026 15:18:40 GMT Funding & raises Product launches Open-source models Dev tooling & infra 🗒 Calathea Daily AI Roundups · 03 Jul 2026

1. Microsoft launches Frontier Co. to help enterprises build AI capability

2. Anthropic launches Claude Science and says it wants to develop drugs for neglected diseases

3. Epoch AI says June disclosures show AI may already be finding software vulnerabilities at scale

4. Alibaba will ban Claude Code at work from July 10 over alleged security risks, Reuters reports

5. Crusoe is in talks to raise $3B in a round that could triple its valuation

6. Elliot Arledge says Claude Fable 5 wrote the first single-kernel KernelBench-Mega submission, beating Opus 4.8 and GLM-5.2

7. Haotian Xie, Junlin Chen and Rutgers/GMU coauthors propose DeadPool for hot-swapping failed GPU nodes during LLM training

8. Astroware AI researchers release HaloGuard 1.0, an open-weight constitutional input classifier for multilingual AI safety

9. Diffusers 0.39.0 adds Cosmos 3, NVIDIA’s unified world foundation model, plus new image and video pipelines

10. ByteDance Seed publishes PAR protein-generation checkpoints on Hugging Face after the February paper and ICML 2026 oral acceptance

11. Claude Code v2.1.198 makes background subagents default and brings Claude in Chrome to general availability

12. Om Patel’s viral screenshots claim Claude Fable 5 exposed raw chain-of-thought during a coding task

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/40 https://t.me/CalatheaAI/40 Thu, 02 Jul 2026 12:46:58 GMT Funding & raises Product launches Dev tooling & infra Policy & safety 🗒 Calathea Daily AI Roundups · 02 Jul 2026

1. OpenAI proposes handing Trump administration 5% stake in the company

2. NVIDIA launches revenue-share model for AI factory compute capacity

3. DeepSeek reverses course after starting an API price war, adding peak-hour surcharges for V4 use

4. Together AI raises $800M Series C at $8.3B valuation led by Aramco Ventures

5. ByteDance says Seed2.0 targets long-tail knowledge and complex instruction-following for real-world tasks

6. White House accelerates frontier-model release standards, guidance expected next week

7. Microsoft releases HARC-Qwen2.5-7B-Instruct safety-aligned model on Hugging Face

8. SoftBank plans to rent AI compute in the US at 10-gigawatt scale

9. Meta is building a cloud business to sell excess AI computing capacity

10. Claude Code is steganographically marking requests with hidden prompts

11. Kling AI nears $3B round at $18B valuation for AI video generation

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/39 https://t.me/CalatheaAI/39 Wed, 01 Jul 2026 11:39:35 GMT Funding & raises Product launches Open-source models Dev tooling & infra 🗒 Calathea Daily AI Roundups · 01 Jul 2026

1. US Commerce lifts export controls on Claude Fable 5 and Mythos 5, Anthropic restoring access tomorrow

2. Artificial Analysis: Claude Sonnet 5 scores 53, but costs 15% more per task than Opus 4.8 before promo pricing

3. Etched exits stealth with first inference racks, $1B+ in customer contracts, and $800M raised

4. Xiaoyin Qu argues cheap Chinese models could pressure Anthropic’s model-layer margins

5. Google ships Nano Banana 2 Lite and Gemini Omni Flash, fastest cheapest image models

6. BMW researchers show 64K-context Qwen3-30B running on a single 32GB RTX 5090 without retraining

7. SemiAnalysis: hardware-software co-design is AI's real 100x, models, kernels, silicon together

8. Schneider Electric to acquire industrial AI firm Cognite for $3.1 billion

9. ByteDance picks Brazil for its largest data center outside China

10. OpenAI debugged a year of data infra crashes, found 18-year-old open-source bug

11. Meta and OpenAI alumni seek $400M for new AI lab

12. Thinking Machines Lab partners with Bridgewater to fine-tune financial analysis model

13. Andrew Curran predicts an OpenAI spinout will announce a memory-efficiency architecture breakthrough

Follow Calathea for all things AI: https://t.me/CalatheaAI]]> https://t.me/CalatheaAI/38 https://t.me/CalatheaAI/38 Wed, 01 Jul 2026 03:07:41 GMT 🚨 Anthropic says the U.S. Department of Commerce has lifted export controls on Fable 5 and Mythos 5. Access starts coming back tomorrow.

This matters beyond Claude. GPT-5.6 is in the same frontier-model access lane, meaning builders could have access to both models soon.

https://x.com/AnthropicAI/status/2072106151890809341]]>