Cycle on Jun 7, 2026

I will start by listing the active threads and watches for this topic using the `list_threads` tool to see if there are any specific user-prioritized questions or active watches we need to address or check.

I will run a web search to look for the latest AI news and releases from the past week, focusing on major model and product releases, big tech announcements, and any developments related to our active watches (Apple WWDC 2026, SpaceX/xAI IPO, Meta Muse Spark, and Anthropic/OpenAI).

The search results show several highly specific, fresh, and relevant stories from the past few days (early June 2026): 1. **Anthropic's Claude Outage and Notion Rerouting**: On June 5, 2026, Claude services went down, causing Notion to disable all Claude models and reroute to alternatives. 2. **Anthropic's "Mythos Preview" model and Project Glasswing**: India gaining access to Anthropic's advanced AI cybersecurity model (Claude Mythos Preview) under "Project Glasswing". Let's investigate this. It seems connected to the "When AI builds itself" / recursive self-improvement warning. 3. **xAI Distillation of Claude**: xAI reportedly trained its coding models on Claude outputs for months before being cut off. 4. **Trump Proposes Public Fund / Trump's AI Policy**: "Trump Proposes Public Fund to Distribute AI Shares to Americans" (voluntary contributions from AI companies like OpenAI, Anthropic, SpaceX). 5. **Meta Muse Spark Delay**: "Meta Keeps Delaying the Release of Its New AI Model to Developers" (WSJ article from June 4, Reuters article). Let's run a search for these specific stories to fetch and read their details. I'll search for "Project Glasswing" and "Claude Mythos Preview" first, as well as the WSJ Meta Muse Spark article, and the xAI Claude distillation story. Let's start with a search on Project Glasswing and Anthropic Mythos.

Wow, this is a massive story. Anthropic has expanded "Project Glasswing" to 150 more organizations across 15 countries, giving them access to **Claude Mythos Preview**, which is described as a highly capable and secure cybersecurity model. Let's fetch the Anthropic announcements: `https://www.anthropic.com/news/expanding-project-glasswing` and `https://techcrunch.com/2026/06/02/anthropic-scales-claude-mythos-to-critical-infrastructure-in-15-countries/`. Let's do that first to get the exact quotes.

This is amazing and highly specific. Anthropic's "Project Glasswing" was first announced in early April, but on June 2, 2026, they officially expanded it, bringing in 150 more organizations across 15 countries to use **Claude Mythos Preview** (a powerful, cybersecurity-focused model with extraordinary defensive and offensive capabilities). The model has already found over 10,000 high- or critical-severity security flaws in early testing. Let's do a quick search on the other topics: - "xAI trained its coding models on Claude outputs" - "Trump proposes public fund to distribute AI shares" - "Meta keeps delaying the release of its new AI model" (Muse Spark) - "Anthropic Claude Outage Notion" Let's search for xAI training on Claude outputs. This is a fascinating story.

Wow! Look at these massive stories from the last 24 hours: 1. **xAI Distillation of Claude**: *The Information* published an article titled "How xAI Went From Chasing Anthropic to Powering It" on June 6, 2026. It reveals that xAI spent months distilling Anthropic's Claude to train its own coding models, even using personal accounts and Blackbox AI after Anthropic revoked access. 2. **Apollo and Blackstone $35 Billion Chip Deal**: Apollo Global Management and Blackstone finalized a $35 billion debt financing package to buy Google TPU chips for Anthropic! 3. **OpenAI Lockdown Mode**: OpenAI launched "Lockdown Mode" for ChatGPT to block prompt injection data exfiltration. Let's fetch the details of the xAI distillation story. I'll search for this article and fetch some of these pages. Let's start with `https://the-decoder.com/elon-musks-xai-reportedly-trained-its-coding-models-on-claude-outputs-for-months-before-getting-cut-off/` and search for the Apollo and Blackstone deal. Let's search for "Apollo Blackstone 35 billion Anthropic".

This is an incredible development in AI compute and capex! Under this deal, Apollo and Blackstone have finalized a **$35 billion debt/private credit package** for Anthropic to expand its AI infrastructure. The financing will fund the lease of Google's custom TPU chips, and Broadcom is backstopping part of the senior debt! This is a massive, record-shattering private credit deal for AI. Let's fetch the details of this deal first. I'll read the Bloomberg article via a financial or tech site or Yahoo Finance. Let's fetch `https://ca.investing.com/news/company-news/apollo-blackstone-finalize-35-billion-ai-chip-financing-for-anthropic-4678668` or `https://cryptobriefing.com/apollo-blackstone-finalize-35b-funding-for-anthropics-ai-infrastructure-growth/` or yahoo finance. Let's do `https://ca.investing.com/news/company-news/apollo-blackstone-finalize-35-billion-ai-chip-financing-for-anthropic-4678668`. Let's also fetch the xAI distillation article `https://the-decoder.com/elon-musks-xai-reportedly-trained-its-coding-models-on-claude-outputs-for-months-before-getting-cut-off/`. Let's fetch both.

This is incredibly detailed and contains precise financial structures. - **Total Financing**: $35 billion financing package finalized by Apollo Global Management and Blackstone for Anthropic. - **Tranches**: - $6 billion of A1 notes (sold to banks at 1 percentage point over Treasuries, supported by Broadcom). - $24 billion of A2 notes (carried a 5.75% coupon, supported by Broadcom). - $4.5 billion B tranche (priced at 8.5% coupon, does not benefit from Broadcom's support). - **Structure**: Special-purpose vehicle (SPV) purchases Google's custom TPUs using debt and equity, then leases them to Anthropic. - **Broadcom's Role**: Backstops the senior portions (A1 and A2 notes, total $30B) through a "residual value agreement" (covering any shortfall if Anthropic defaults and chips are sold for less than owed). - **Platform**: Part of the "AI XPV platform" aiming to deploy over 20 gigawatts of compute capacity through 2028, mentioned by Broadcom CEO Hock Tan. - **Context**: Follows Anthropic's confidential IPO filing and a funding round valuing the company at $965 billion! Let's now fetch the details of the xAI distillation story. Let's send a fetch for `https://the-decoder.com/elon-musks-xai-reportedly-trained-its-coding-models-on-claude-outputs-for-months-before-getting-cut-off/`.

This is a massive and fascinating development: - **xAI and Claude Distillation**: xAI spent months distilling Anthropic's Claude models to train its own coding models. This direct training on Claude's outputs continued even after Anthropic revoked official access in January 2026. xAI engineers bypassed this by using personal accounts and an intermediary service called Blackbox AI. - **Internal Turbulence at xAI**: - Pretraining team shrank to under five people. - Four Grok code leads left within months, along with many co-founders (due to safety concerns and frustration over Grok's failure to catch up). - An employee accidentally deleted critical training data, delaying work by two to three weeks. - **Compute Rental**: Musk is renting out his massive Colossus compute infrastructure to Anthropic (via SpaceX, for 220,000 GPUs) and to Google (via SpaceX, a $920 million/month deal for 110,000 Nvidia AI chips) instead of training his own models with it right now. Let's search for "The Information" article "How xAI Went From Chasing Anthropic to Powering It" or look for more details to see if we can read the original article. It seems we have a lot of detailed points already, but let's do a search on `How xAI Went From Chasing Anthropic to Powering It` to see if there are other details.

These search results from *The Information* are incredibly rich with details that perfectly update our understanding of the frontier AI space: 1. **SpaceX IPO and Financial Forecasts**: - SpaceX's IPO is planned for next week (June 12, 2026), targeting the "largest initial public offering of all time." - Wall Street forecasts that SpaceX will burn **$350 billion of cash through 2030**! - SpaceX's compute rental deal with Anthropic runs for a minimum of six months, according to an amended securities filing on Monday (June 1, 2026), providing an immediate revenue stream. - The deal flags data center water risks. 2. **Anthropic IPO and Valuation Details**: - Anthropic filed a draft S-1 form to the SEC for a confidential IPO. - The confidential IPO options/filing came shortly after a funding round that valued Anthropic at **$965 billion** (close to $1 trillion!). - Salesforce's stake in Anthropic has reached roughly **$5 billion** following multiple investments. 3. **xAI Distillation of Claude**: - xAI used Claude models for distillation and training, including personal accounts and Blackbox AI after being cut off in January 2026. - xAI went from chasing Anthropic to renting a major portion of its Colossus compute infrastructure to them (via SpaceX's 220,000 GPUs deal). Let's do a search on the "SpaceX Forecasted to Burn $350B by 2030" and "SpaceX Discloses Updated Details of Anthropic Deal" to fetch and read these. Let's search for "SpaceX Discloses Updated Details of Anthropic Deal" first.

This is an incredibly rich set of details from the WSJ, The Information, and Substack regarding the upcoming SpaceX/xAI IPO on June 12, 2026! Let's fetch `https://michaelparekh.substack.com/p/ai-google-and-anthropic-help-paint` to read the full context. Substack posts are usually free to fetch and contain great syntheses of paywalled articles like WSJ and The Information. Let's fetch it.

This is brilliant and extremely detailed. Let's summarize the key pieces of information from this Substack article (which references the Wall Street Journal and other recent news from June 2026): 1. **SpaceX/xAI IPO**: - The IPO is scheduled for **June 12, 2026**, targeting a valuation of **$1.75 - $1.77 trillion** (with SpaceX/xAI merged or presented together, trading on the Nasdaq). - Expected to raise **$75+ billion** in proceeds, with over $30 billion coming from individual retail investors. - Wall Street forecasts SpaceX will burn $350 billion through 2030, but the company is dressing up its S-1 prospectus with massive short-term compute rental deals to rivals. 2. **Google Cloud's $30 Billion SpaceX Compute Deal**: - Google has agreed to pay SpaceX **$920 million a month** from October 2026 to June 2029 to rent data-center capacity in the Colossus complex (built by xAI). - The deal includes the compute capacity of at least **110,000 Nvidia chips**. - The deal ramps up over the summer. Google has the right to cancel in October if chips aren't provided. Otherwise, either party can cancel starting next year with **90 days' notice**. - This is a bridge capacity deal for Google's Gemini Enterprise agent platform. 3. **Anthropic's SpaceX Compute Deal Details**: - Anthropic's deal is even larger: they agreed to pay SpaceX **$1.25 billion a month** to rent 220,000 Nvidia chips in the Colossus data center. - However, an amended securities filing on Monday, June 1, 2026, revealed that this deal runs for a **minimum of only six months** (180 days), after which it is cancelable. It was pitched as a massive multi-year deal initially, but the fine print shows a 180-day minimum. 4. **Orbital Data Centers (Space Data Centers)**: - SpaceX has pitched plans to build data centers in space. - Google is working with Planet Labs on its own orbital data centers by 2027 under a moonshot project called **Project Suncatcher**. - Musk said SpaceX's future orbital data centers will run chips from multiple companies, including Google's custom TPUs. Anthropic also expressed interest in orbital data centers if they launch. Let's double check if there are any updates regarding our other active watches: - **Meta Muse Spark API**: Let's check if there is an update on Meta Muse Spark. The search results showed "Meta Keeps Delaying the Release of Its New AI Model to Developers" (WSJ, June 4, 2026). Let's search for "Meta Muse Spark June 2026" or "Meta Muse Spark delay" to see what's happening. - **Apple WWDC 2026**: WWDC starts on June 9, 2026 (two days from now). We should confirm the current expectations and see if any last-minute leaks or details have emerged. Let's do a search on Meta Muse Spark delay first.

This is a perfect update for our active watch on **Meta Muse Spark Developer API Release** (thread id: `019e9889-2962-79c0-ba91-3aa809f9f6b9`). According to the WSJ reporting from June 3-4, 2026: - Meta has repeatedly pushed back plans to release its new Muse Spark AI model API to developers. - As of Tuesday (June 2, 2026), there was no scheduled or set launch date. - The delays are due to bugs and infrastructure issues, presenting a setback for Meta's ability to monetize its massive AI investments. - Meta is still testing with select early partners (like Alexandr Wang's Scale AI or other early partners), and some sources claim they still hope to launch "later this month" (June 2026), but there is currently no set date. - This is a direct update on the watch we registered. We can record an event for this, but since the watch is for June 30, 2026, and the month has not passed yet, we can record a partial update (outcome) or keep tracking. Let's record an ongoing event or write a note about this delay. Let's look into **Apple's WWDC 2026** (starts June 9, 2026). Let's search for "Apple WWDC 2026 Siri iOS 27 AI agent" or similar to see the latest expectations.

We have a lot of great information here. First, WWDC 2026 starts on **Monday, June 8, 2026** (or June 9 in some timezones, but June 8 is the keynote). So the watch `019e9bc8-00dd-7d1d-b044-b9c333710a79` has its `fire_date` set to June 9, 2026, which is exactly WWDC. This watch is a "dated catalyst" that will fire when WWDC happens and the official announcements are made. Since today is Sunday, June 7, 2026, the WWDC keynote has not happened yet (it happens tomorrow, Monday, June 8, 2026). So we should keep this watch open and check it next cycle, but we can write a note summarizing the absolute latest expectations and leaks for WWDC 2026. Wait, let's look at the active watches again: 1. `019e9bc8-00dd-7d1d-b044-b9c333710a79` — Apple WWDC 2026 Siri and iOS 27 AI Agent Releases (fire_date: 2026-06-09T00:00:00Z) 2. `019e9889-2962-79c0-ba91-3aa809f9f6b9` — Meta Muse Spark Developer API Release (fire_date: 2026-06-30T00:00:00Z) 3. `019e9889-19fc-7370-b885-054439ee4234` — SpaceX and xAI Public Debut / IPO (fire_date: 2026-06-12T00:00:00Z) None of these watches have fully resolved yet (as today is June 7, 2026, ahead of June 8/9, June 12, and June 30). However, we have extremely important updates on all three! - **SpaceX/xAI IPO**: The S-1 filings and recent news (WSJ, The Information, Substack) have disclosed massive details: the $1.75T - $1.77T valuation, the $75+ billion expected proceeds, the $350 billion cash burn forecast by 2030, and the multi-billion dollar monthly compute rental deals with Google ($920M/month) and Anthropic ($1.25B/month). - **Meta Muse Spark API**: Pushed back repeatedly due to bugs and infrastructure, with no set release date as of early June, although Meta claims they are still targeting June. - **Apple WWDC 2026**: High expectations for Gemini-powered Siri, iOS 27, and Apple Intelligence. Wait, let's write findings for these! Let's see: what else is new? - **Anthropic Project Glasswing expansion**: This is a major update to the prior finding `anthropic-recursive-self-improvement-pause-call` or a new finding. Let's look at the prior findings: - `anthropic-recursive-self-improvement-pause-call`: In a landmark paper on June 4, 2026, titled "When AI builds itself," Anthropic revealed Claude models reached a pivotal milestone in autonomous software... - Wait, Project Glasswing was expanded on **June 2, 2026**, giving 150 more organizations across 15 countries access to **Claude Mythos Preview** (the powerful cybersecurity-focused model). This model has already found over 10,000 high- or critical-severity flaws. This is directly related! We can write a specific finding for this, or update `anthropic-recursive-self-improvement-pause-call`. Let's write a fresh finding for Project Glasswing or update the existing one. Actually, "Project Glasswing" is a specific cybersecurity initiative, and "Claude Mythos Preview" is the model behind it. Let's write a fresh finding for the Glasswing expansion and Mythos model, as it's a major development in the cybersecurity/frontier space. - **Apollo and Blackstone $35 Billion Chip Deal**: This is a huge capex/compute story! Apollo and Blackstone finalized a record-shattering $35 billion debt financing package for Anthropic to lease Google TPUs, with Broadcom backstopping the senior debt. This is a massive new finding. Let's write a note about this. - **xAI Distilling Claude / Internal Troubles**: xAI trained its coding models on Claude outputs for months, even after being cut off in January, bypassing the ban with personal accounts and Blackbox AI. Meanwhile, xAI faces internal turmoil (pretraining team < 5 people, code leads leaving, data deletion). This is a massive new finding. Let's write a note about this. - **SpaceX & Google Compute Deal / S-1 Disclosures**: Google signs a $920M/month deal ($30B total) to rent 110,000 Nvidia chips from SpaceX's Colossus complex. This, along with Anthropic's $1.25B/month deal, helps dress up SpaceX's S-1 prospectus ahead of its $1.77T IPO on June 12, 2026. This is a massive new finding. Let's write a note about this. - **Meta Muse Spark Delay**: Meta has repeatedly delayed the developer API for Muse Spark due to bugs and infrastructure, with no set release date as of early June. This is a massive new finding. Let's write a note about this. Let's double-check the exact instructions: "For each distinct theme or story worth persisting, call write_note — ONE note per theme, never a combined write-up. A finding TELLS THE STORY THROUGH CONCRETE EVIDENCE — it does not just describe the vibe. Every substantive finding must: - name and link the actual thing (the story, article, paper, or page that was posted/published) — not just "the discussion", - carry 1-2 REPRESENTATIVE VERBATIM QUOTES, copied exactly, each kept with the link it came from (an HN comment permalink, the article URL, etc.), - then say what it means." Let's write these findings. Let's prepare the first finding: **Apollo and Blackstone $35 Billion debt financing package for Anthropic**. Let's review the details from `https://ca.investing.com/news/company-news/apollo-blackstone-finalize-35-billion-ai-chip-financing-for-anthropic-4678668`. - Title: Apollo and Blackstone Finalize Record $35 Billion Debt Deal to Fund Anthropic's AI Compute - Slug: apollo-blackstone-35b-anthropic-compute-financing - Body: We will quote Simon Mugo's Investing.com report: > "Apollo Global Management and Blackstone have finalized a $35 billion financing package for Anthropic, providing funding for artificial intelligence infrastructure in one of the largest private credit transactions on record..." And: > "The senior portions of the financing include $6 billion of A1 notes and $24 billion of A2 notes that are supported by Broadcom through a residual value agreement. The arrangement helped the notes secure investment-grade private ratings and lower borrowing costs." Let's write this note. I will formulate the markdown carefully. No YAML frontmatter, first heading is `# `, slug is `apollo-blackstone-35b-anthropic-compute-financing`. Let's call `write_note`.

Now let's write the second finding: **xAI's Distillation of Anthropic's Claude and Internal Turmoil**. Let's review the details from `https://the-decoder.com/elon-musks-xai-reportedly-trained-its-coding-models-on-claude-outputs-for-months-before-getting-cut-off/` and *The Information*. - Title: xAI Secretly Spent Months Distilling Claude to Train Grok Coding Models Amid Internal Turmoil - Slug: xai-distilled-claude-internal-turmoil - Body: We will quote: > "Elon Musk's xAI spent months distilling Anthropic's Claude to train its own coding models, The Information reports. xAI trained its coding model directly on Claude's outputs. After Anthropic revoked official access in January, xAI engineers kept going through personal accounts and the intermediary service Blackbox AI." And: > "Internally, xAI looks troubled. The pretraining team shrank to under five people. Four Grok code leads left within months, along with many co-founders. One employee accidentally deleted critical training data, costing two to three weeks of work, according to The Information." Let's write this note. Let's call `write_note`.

Now let's write the third finding: **SpaceX/xAI IPO S-1 Disclosures & Multi-Billion Dollar Compute Rental Deals (Google & Anthropic)**. Let's review the details from `https://michaelparekh.substack.com/p/ai-google-and-anthropic-help-paint` (WSJ reporting). - Title: SpaceX Preps for $1.77 Trillion IPO with Massive Compute Rental Deals for Google and Anthropic - Slug: spacex-google-anthropic-compute-deals-ipo-preps - Body: We will quote: > "Google will pay SpaceX $920 million a month from October 2026 to June 2029 in a deal that includes the computing capacity of at least 110,000 Nvidia chips, according to a SpaceX securities filing on Friday." And: > "In its IPO prospectus, SpaceX said that Anthropic agreed to pay $1.25 billion a month for compute capacity in the Colossus data center built by xAI, in a deal that runs through 2029." And we will note: - Google's deal is a bridge capacity deal for its Gemini Enterprise agent platform. - Anthropic's deal is for 220,000 GPUs, but an amended securities filing on June 1, 2026, revealed that Anthropic's deal has a minimum commitment of only six months (180 days), after which it is cancelable. - The IPO is set for June 12, 2026, targeting a valuation of $1.75 - $1.77 trillion, raising $75+ billion. - Wall Street forecasts that SpaceX will burn $350 billion of cash through 2030, which explains the aggressive search for high-margin AI infrastructure revenue. Let's write this note. Let's call `write_note`.

Now let's write the fourth finding: **Anthropic's Project Glasswing Expansion & Claude Mythos Preview**. Let's review the details from `https://www.anthropic.com/news/expanding-project-glasswing` and `https://techcrunch.com/2026/06/02/anthropic-scales-claude-mythos-to-critical-infrastructure-in-15-countries/`. - Title: Anthropic Expands Project Glasswing to 150 Organizations to Deploy Cybersecurity-Focused Claude Mythos Preview - Slug: anthropic-expands-project-glasswing-claude-mythos - Body: We will quote: > "We recently described how these partners have so far found more than 10,000 high- or critical-severity security flaws... We’re now expanding Project Glasswing... extending the partnership to approximately 150 new organizations... based in more than 15 countries..." And: > "Mythos Preview continues a long-term trend that we’ve been warning about for some time: within 6 to 12 months, we expect that many other AI companies will have Mythos-class models, and they could release them without safeguards that prevent misuse." And we will note: - Project Glasswing is a collaborative program to secure critical software infrastructure. - Initial 50 partners found over 10,000 critical security flaws. - The model behind this is **Claude Mythos Preview**, which has extraordinary cybersecurity capabilities (both defensive and offensive). - Anthropic is warning about the imminent arrival of "Mythos-class models" from competitors without safeguards. - This connects directly to Anthropic's warnings about recursive self-improvement and AI safety (see [[anthropic-recursive-self-improvement-pause-call]]). Let's write this note. Let's call `write_note`.

Now let's write the fifth finding: **Meta's Repeated Muse Spark API Delays due to Bugs and Infrastructure**. Let's review the details from WSJ/Reuters reporting on June 3-4, 2026. - Title: Meta Delays Muse Spark API Launch for Developers Over Bugs and Infrastructure Issues - Slug: meta-delays-muse-spark-api-bugs - Body: We will quote: > "Meta has repeatedly pushed back plans to release its new Muse Spark AI model API to developers, and as of Tuesday, had no scheduled launch date, according to a Wall Street Journal report." And: > "The shifting timeline is a setback for Meta's ability to monetize its massive investments in building frontier AI models." And we will note: - Muse Spark is Meta's closed-source, high-capability model that powers its main AI features. - The API was originally promised "soon" in April, but has faced repeated postponements due to bugs and infrastructure challenges. - While Meta claims testing with early partners continues and they still hope to launch in June 2026, the lack of a firm release date highlights the operational bottlenecks Meta faces in commercializing its frontier models. Let's write this note. Let's call `write_note`.

Now let's check if there are other threads or watches we should update or resolve. Let's review the active threads: 1. `019e9bc8-00dd-7d1d-b044-b9c333710a79` (Apple WWDC 2026 Siri and iOS 27 AI Agent Releases) — This watch is for WWDC, which starts tomorrow, June 8, 2026. Since the keynote hasn't happened yet, we cannot resolve this thread. We will leave it open so the next cycle (which will run after WWDC) can resolve it. 2. `019e9889-2962-79c0-ba91-3aa809f9f6b9` (Meta Muse Spark Developer API Release) — This watch tracks whether the API successfully ships in June 2026 or faces further delays. Since it's currently June 7, 2026, and the API has not shipped yet (and is delayed), we should not resolve it yet. We will leave it open to see if Meta manages to release it by June 30, 2026. 3. `019e9889-19fc-7370-b885-054439ee4234` (SpaceX and xAI Public Debut / IPO) — This watch fires when the IPO takes place on June 12, 2026. Since today is June 7, we should leave it open and let the next cycle resolve it once the IPO has occurred. Wait, are there any other threads we should open or record? Let's look at the Apollo/Blackstone $35B deal or xAI's internal turmoil. We have written very solid notes for both. Wait! Let's check if there are any other interesting stories we found in our searches: - **OpenAI Lockdown Mode**: "OpenAI launches Lockdown Mode for ChatGPT to block prompt injection data exfiltration by disabling web browsing, deep research, and agent mode." This is a security feature. Let's do a quick search on this to see if it's worth a finding. Let's search "OpenAI Lockdown Mode ChatGPT".

This is another fantastic, highly specific story about a major product release from OpenAI this week! Let's write a note about this. Let's review the details: - Title: OpenAI Launches ChatGPT "Lockdown Mode" to Prevent Prompt Injection Data Exfiltration - Slug: openai-chatgpt-lockdown-mode - Body: We will quote: > "Even with Lockdown Mode, ChatGPT could be still vulnerable to prompt injections, but the goal is to reduce the likelihood that sensitive data gets shared in..." (from TechCrunch) And: > "When enabled, Lockdown Mode disables live web browsing, Deep Research, Agent Mode, and limits external connections to protect against potential attacks." And Simon Willison's explanation: - It is designed to help prevent the final stage of data exfiltration from a prompt injection attack by limiting outbound network requests that could transmit stolen data. - Rolling out to Free, Plus, Pro, and self-serve ChatGPT Business accounts. Let's write this note. Let's call `write_note`.

This is amazing. We have written 5 highly detailed and extremely fresh notes covering the absolute frontier of AI: 1. `apollo-blackstone-35b-anthropic-compute-financing` (Apollo/Blackstone $35B debt financing for Anthropic) 2. `xai-distilled-claude-internal-turmoil` (xAI's secret distillation of Claude and internal turmoil) 3. `spacex-google-anthropic-compute-deals-ipo-preps` (SpaceX S-1 disclosures of compute rental deals with Google and Anthropic ahead of $1.77T IPO) 4. `anthropic-expands-project-glasswing-claude-mythos` (Anthropic's Project Glasswing expansion and Claude Mythos Preview) 5. `meta-delays-muse-spark-api-bugs` (Meta's repeated delays of the Muse Spark developer API due to bugs) 6. `openai-chatgpt-lockdown-mode` (OpenAI's new ChatGPT Lockdown Mode) Wait, this is 6 notes! Let's double check if there are any other actions to take. - "Call finish_cycle to end the cycle." - "This is the ONLY way to complete a cycle — do not just stop after researching." Let's call `finish_cycle`. I will not pass `no_findings_reason` because we wrote 6 excellent findings this cycle. Let's finish the cycle!

I have completed the research cycle and successfully persisted six highly relevant, detailed, and fresh findings on the active AI frontier. Here is a summary of the key developments that shipped or were disclosed this week: ### 1. **Apollo & Blackstone Finalize Record $35 Billion Debt Package for Anthropic** In one of the largest private credit deals on record, Apollo and Blackstone have finalized a **$35 billion debt financing package** for Anthropic. This deal utilizes a special-purpose vehicle (SPV) to purchase Google’s custom Tensor Processing Units (TPUs) and lease them to Anthropic. Crucially, **Broadcom** is backstopping $30 billion of the senior debt (Tranches A1 and A2) through a residual value agreement, securing investment-grade private ratings and significantly lowering Anthropic's borrowing costs. This represents a massive shift in AI capex, leveraging structured private credit to build out massive infrastructure without diluting equity. *(Persisted as: `[[apollo-blackstone-35b-anthropic-compute-financing]]`)* ### 2. **SpaceX Discloses Multi-Billion Dollar Compute Rentals Ahead of $1.77 Trillion IPO** Ahead of its highly anticipated Nasdaq IPO on **June 12, 2026** (targeting a valuation of $1.75 - $1.77 trillion and raising $75+ billion), SpaceX's S-1 filings disclosed massive, near-term compute rental deals at its Colossus complex: * **Google Cloud** will pay SpaceX **$920 million a month** from October 2026 to June 2029 to lease at least 110,000 Nvidia chips as "bridge capacity" for its surging Gemini Enterprise agent platform. * **Anthropic** agreed to pay **$1.25 billion a month** to lease 220,000 chips. However, an amended filing on June 1 revealed this contract has a minimum commitment of **only six months (180 days)**, after which it is cancelable. These short-term, high-margin deals provide immediate revenue to offset Wall Street's forecasts that SpaceX will burn **$350 billion of cash through 2030**. *(Persisted as: `[[spacex-google-anthropic-compute-deals-ipo-preps]]`)* ### 3. **xAI Secretly Distilled Claude Amid Severe Internal Instability** *The Information* revealed that Elon Musk's xAI spent months distilling Anthropic's Claude outputs to train its own coding models. Even after Anthropic discovered this and revoked xAI's API access in January, xAI engineers bypassed the ban using personal accounts and the intermediary service Blackbox AI. This reliance on rival models comes alongside reports of severe internal turmoil at xAI, including a pretraining team that has shrunk to under five people, the departure of four separate Grok code leads, and a major blunder where an employee accidentally deleted a critical pretraining data batch, setting development back by two to three weeks. *(Persisted as: `[[xai-distilled-claude-internal-turmoil]]`)* ### 4. **Anthropic Expands Project Glasswing and Claude Mythos Preview** On June 2, 2026, Anthropic expanded **Project Glasswing**, extending access to its highly specialized, powerhouse cybersecurity model, **Claude Mythos Preview**, to 150 new organizations across 15 countries. The model is being deployed to secure critical software infrastructure (power, water, healthcare, communications, and hardware). In initial testing, Glasswing partners used Mythos to scan codebases and identified over **10,000 high- or critical-severity security flaws** in weeks. Anthropic warns that similar "Mythos-class" systems will be developed by competitors within 6 to 12 months, raising the stakes for defensive preparedness. *(Persisted as: `[[anthropic-expands-project-glasswing-claude-mythos]]`)* ### 5. **Meta Delays Muse Spark Developer API Over Bugs and Infrastructure Issues** According to reports from *The Wall Street Journal*, Meta has repeatedly postponed the developer API release for its closed-source, high-capability **Muse Spark** model. Originally promised "soon" in April, the API remains in limbo with no scheduled launch date as of early June due to persistent bugs and infrastructure scaling bottlenecks. While Meta claims testing with select partners continues and they still target a June release, the ongoing delays slow down Meta's ability to monetize its massive AI investments and build out a competitive developer ecosystem. *(Persisted as: `[[meta-delays-muse-spark-api-bugs]]`)* ### 6. **OpenAI Launches ChatGPT "Lockdown Mode" Against Data Exfiltration** OpenAI began rolling out **Lockdown Mode**, an optional security setting designed to protect ChatGPT users from data exfiltration caused by prompt injection attacks. Rather than trying to block the injection itself, Lockdown Mode sandbox-restricts the session by disabling live web browsing, Deep Research, and Agent Mode, and heavily limiting external network requests. This feature is rolling out to Free, Plus, Pro, and self-serve ChatGPT Business accounts, allowing users to trade advanced capabilities for enterprise-grade data isolation. *(Persisted as: `[[openai-chatgpt-lockdown-mode]]`)* *** ### Active Watches Status: * **Apple WWDC 2026 Siri and iOS 27 AI Agent Releases** (Thread ID: `019e9bc8-00dd-7d1d-b044-b9c333710a79`): The keynote starts tomorrow, **Monday, June 8, 2026**. Leaks point to a major Gemini-enhanced Siri rebuild, generative Photos tools in iOS 27, and "Apple Intelligence" updates. This watch will remain open to be resolved in the next cycle. * **SpaceX and xAI Public Debut / IPO** (Thread ID: `019e9889-19fc-7370-b885-054439ee4234`): Set for **June 12, 2026**. This watch remains open to track incoming financial disclosures and debut trading next week. * **Meta Muse Spark Developer API Release** (Thread ID: `019e9889-2962-79c0-ba91-3aa809f9f6b9`): Delayed but still targeted for late June. This watch remains open to trace whether Meta delivers by June 30 or faces further setbacks.