X Freeze (@XFreeze)

SpaceX just laid out a major plan to protect and restore the coastline around its launch site Boca Chica Beach is staying public as Starbase keeps growing More than 50 tons of trash already removed through weekly beach cleanups with the City of Starbase Tens of millions of dollars offered to fight erosion, restore the shoreline and keep public access open Ongoing support for conservation groups rescuing sea turtles The proposed land purchase covers about 1,200 acres around the launch facility, with Boca Chica Beach and the Texas coastline left out of it Around 600 of those acres would be permanently protected as sensitive coastal habitat Money from the sale would help Texas build new parks and recreation across the Rio Grande Valley Starbase grows, and the beach stays open to everyone

Elon Musk just revealed a massive update on Grok 4.8, and SpaceXSI has completely rebuilt its reinforcement learning and inference infrastructure from scratch The engineering team has abandoned JAX and SGLang, rewriting the entire RL and inference stack in C/C++ Here’s where Grok 4.8 stands: • Pre-training is complete • Major cleanup of supervised fine-tuning data • Reinforcement learning now underway • Evaluation scores are starting to climb • Soft rollout to Grok Bot users planned for later this week And Elon also gave another imp update about Grok 4.8: “Grok 4.8 will bear no resemblance to 4.7 (thankfully)” He also explained why the development took longer than expected: “It is not yet incredible, but, this time, the foundations are built right, so I expect it to evolve to be something special” This is a fundamental engineering overhaul, not just another model upgrade SpaceXSI has rewritten the underlying infrastructure to build a much stronger foundation for Grok’s future development. And the first Grok Bot users are expected to get their hands on it this week

Elon Musk on why internet access is one of the most powerful tools humanity has for lifting people out of poverty: “The single biggest thing you can do to lift people out of poverty and help them is giving them an internet connection” Because once someone gets connected, their opportunities completely change • they can learn almost anything for free • get access education and information • Find work and build businesses • Sell goods and services to the global market • Participate in an economy far beyond their local community That is the bigger mission behind Starlink Bringing low-cost, high-bandwidth internet to places where traditional connectivity is unavailable, too expensive or simply never reaches For hundreds of millions of people, an internet connection is not just a convenience....it is access to new opportunity they never had before

SpaceX just got a massive FCC green light for Starlink Mobile The FCC has approved SpaceX to launch and operate up to 15,000 next-generation satellites built for direct-to-phone connectivity The new system is designed to: • Target peak speeds of up to 150 Mbps per user • Connect ordinary cell phones directly from space • Operate in very low orbit at roughly 326–335 km • Use T-Mobile spectrum alongside spectrum SpaceX is acquiring from EchoStar • Deliver orders of magnitude more capacity than the first-generation Starlink Mobile constellation SpaceX currently has roughly 650 dedicated mobile satellites in orbit The FCC also granted a waiver giving SpaceX a path to provide satellite wireless service without requiring a spectrum-leasing agreement with a terrestrial mobile operator Starlink Mobile is moving far beyond basic dead-zone connectivity SpaceX is building toward a full 5G cellular network in space

Grok Build just got one of its biggest updates yet.....a massive upgrade touching almost every part of the agent stack Compact mode gets a major upgrade, timestamps are now configurable, shell commands clearly show when they’re running or finished, tables can export directly as TSV, CSV or Markdown, and the dashboard now gives much better visibility and control over worktrees Under the hood, the update is even bigger: stalled model calls now retry, long sessions are more resilient, permission handling is much tighter, failed file writes preserve existing contents, CJK memory search improves, concurrent config edits are safer, MCP tool discovery is more reliable, and worktree creation + deletion gets a huge reliability pass There’s even a breaking safety change: grok worktree rm now refuses to delete active worktrees or ones with unsaved changes unless you explicitly force it Release Notes: v1.0.50 Breaking Changes: • grok worktree rm without -f now refuses to delete worktrees that are in use or contain unsaved changes and exits with an error. Features: • Compact mode now collapses recaps and keeps inner spacing on tall terminals. • Clicking the git branch in the header now copies the full branch name. • Long branch names now abbreviate on narrow terminals while remaining copyable. • Targeted cancels now stop only the specific response when requested by clients. • Compact mode now hides the dock (or tasks pane) the same way Ctrl+G does. • New timestamps setting offers off, minimal, and on modes in Settings and /timestamps. • Shell command output now shows "Running …" while executing and "Ran …" when finished. • Live preview pins now include whether an app was observed behind the URL. • Command palette (Ctrl+P) and model picker now open at the same width as the settings modal. • Tables rendered by grok now show copy buttons underneath that export clean TSV, CSV or Markdown suitable for spreadsheets. • Dashboard now shows which worktree each session belongs to, displays the -m model in the new-agent box, and lets you toggle worktree mode in the folder picker. Bug Fixes: • Shell permission decisions now correctly prompt or approve commands matching the permission manager. • Cancelled turns now correctly mark unanswered tool calls as cancelled instead of succeeded. • Long error messages in the /usage and /session-info modals now wrap instead of being truncated. • Em-dashes no longer appear in status messages and subagent rows. • Stalled model calls during turns now retry before failing the turn. • Queue pane and Ctrl+; now appear and work when the dock is hidden with Ctrl+G. • Long sessions no longer disconnect when the conversation exceeds 4 MiB. • Permission rules now accept native tool names such as run_terminal_command and Shell. • /rewind now places the cursor at the end of the restored prompt. • Failed file writes now leave the previous contents instead of an empty file. • Fixed /loop and scheduler_create so non-ASCII intervals like 5分 no longer crash the session. • Fixed --disallowed-tools when only one scheduler tool is named; the other two no longer cause startup failure. • Fixed closing a busy tab so queued prompts are no longer sent to the model. • Fixed relative --cwd so grok --cwd a no longer ends up running inside a/a. • The terminal title now shows "Action Required" for any background session waiting on approval. • grok now accepts a prompt via piped stdin in headless mode instead of trying to open the UI. • An admin-set [features] title_refresh = false in requirements.toml now correctly disables title refresh even when GROK_TITLE_REFRESH is set. • Custom notification commands defined under [[ui.notifications.hooks]] now execute on matching events. • An admin pin of [cli] use_leader = false now correctly prevents --leader from starting a leader process. • When multiple MCP servers expose a tool with the same name, search_tool now returns all of them. • Concurrent Grok windows editing settings no longer lose or corrupt config.toml entries such as MCP servers or permission rules. • web_fetch no longer downloads huge files into memory before rejecting them. • memory_get no longer returns huge files that exceed the read limit. • Continue to run no longer kills commands running inside background subagents. • memory_search now finds Japanese, Chinese, and Korean words inside longer notes. • Drafts saved by the model now appear under their actual titles in /feedback. • Imagine replies that are not valid images are now rejected instead of being saved as .jpg files. • A workflow file that fails to parse now reports the exact file and parse error instead of 'unknown workflow'. • Re-encoded images are now oriented correctly using their EXIF tags instead of appearing sideways. • Image and log paths reported after /fork --worktree now point to real files instead of missing locations. • Tab characters in terminal command output are now preserved when shown to the model. • Transparent images are now kept as PNG when re-encoded instead of turning black. • The bundled ripgrep is now available to terminal commands even on images that do not install it. • Request bodies now stay under a small configured max_request_bytes limit instead of growing past it. • Effort level now stays the same after a bare /model switch instead of jumping to the new model's default. • Ctrl+\ and other modified digits no longer accidentally select options on permission or cancel cards. • Feedback submissions now show accurate success or failure messages without premature thanks or duplicated prefixes. • grok doctor now reports the exact config file or setting that prevents Grok from starting instead of saying zero issues. • Recap summaries now appear on screen automatically instead of requiring manual scrolling. • grok doctor fix now opens its confirmation card on Cancel to prevent accidental edits to shell or tmux config. • Slash commands now run regardless of case (e.g. /COMPACT or /Compact). • /context no longer shows an invented auto-compact threshold when the agent does not report one. • Agent text no longer keeps the pager redrawing when a tool call interrupts the sentence. • CLI tool calls no longer fail or overflow when the workspace directory is deleted or output is huge. • Logs written to files or when NO_COLOR is set are now plain text instead of containing color codes. • Prompts now fail immediately with a clear message if the session's working directory no longer exists instead of failing later or recreating it. • grok -w now works correctly with separate-git-dir repositories and shows clear errors when run outside a git repo or on an empty repo. • Shell tool descriptions now correctly state the actual output truncation limit, and grep respects globs relative to the current session directory. • Compaction messages now scroll into view instead of appearing below the bottom of the screen after a turn finishes. • Git-mode worktrees created with -w now preserve uncommitted changes instead of dropping them. • Worktrees created with --worktree-ref no longer contain wrong submodule contents or LFS pointer text. • Worktree creation now shows the same clear error messages on every code path. • Failed worktree creates now show a clear error on the home screen instead of being hidden. • Merge and rebase conflicts now appear in the changes list instead of being hidden. • Fixed a crash when starting a new session with durable rewind checkpoints enabled. • Managed permission files with one bad key no longer lose all their valid rules. • Simultaneous worktree creates with the same label now succeed instead of one failing with "already exists". • grok worktree rm without -f no longer deletes arbitrary folders that happen to share a name with a removed worktree. • Discard now affects only the exact files you select instead of also deleting or reverting similarly-named files. • Permission rules with one bad entry no longer lose the other rules in the same section. • Chat Completions streams that end without a finish reason are now retried instead of returning partial output. • grok worktree rm by label now refuses when the label exists in multiple repos. • Grep now respects Read deny rules that use a leading ./. Performance: • Command hooks no longer cause large memory spikes when they print a lot of output. • Retry waits now vary between processes instead of being identical.

Michael Nicolls just just laid out SpaceX's ambition for connectivity that extends beyond Earth “We will connect the Moon with a constellation, and then we will connect Mars” "Our goal, again, is to extend consciousness beyond this planet, and we want to go and connect the Moon and Mars But before that on Starlink, we're creating the technologies to connect the Moon with very high-bandwidth connectivity systems, very high-bandwidth lasers And we will connect the Moon with a constellation, and then we will connect Mars, and we will form an interplanetary mesh of connectivity that is fundamentally driven by the same technologies we're working on with Starlink To make this happen, SpaceX needs to keep improving safety in LEO so everyone can operate and scale safely” Build safe, coordinated infrastructure in LEO first then extend connectivity and human presence to the Moon, Mars, and beyond

Grok 4.7 takes the TOP 3 spots on VulcanBench Frontier v4, outperforming Fable 5.1, Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and other leading frontier models • Grok 4.7 xHigh - 93.15 • Grok 4.7 High - 92.71 • Grok 4.7 Medium - 92.30 VulcanBench Frontier v4 tests whether AI models can handle difficult repository-level software engineering work rather than simply generate isolated code snippets The benchmark contains 23 behavioural-reconstruction tasks where models have to rebuild legacy software to match its actual behaviour, with functional correctness tested through hidden tests alongside security, lint/complexity and code quality And the task results are just as impressive: • Grok 4.7 xHigh - 23/23 passed • Grok 4.7 High - 22/23 passed • Grok 4.7 Medium - 21/23 passed Grok 4.7 took all #1#, #2# and #3# spots

Grok Bot passed Google’s “I’m not a robot” test when it launched 😂 The other agents… wherever the hell they are, I still haven’t seen a single video of them doing this

Grok 4.7 xHigh ranked #1# on Artificial Analysis’ Cyber Index Grok 4.7 is now very powerful with serious cybersecurity capabilities It outperformed Claude Opus 5.5, Fable 5.1, ChatGPT 6 Astra and other frontier models The Cyber Index combines three evaluations: • CWE-Bench-AA • DeepsecBench-AA • CyberGym-E2E-AA Grok 4.7 is now sitting right at the frontier of cybersecurity reasoning

Elon was thinking about superintelligence long before anyone in the mainstream even understood where superintelligence was heading He was deeply involved in the conversation around Nick Bostrom’s Superintelligence and even contributed to the book. Bostrom thanked him by name in the foreword Elon was already warning about the safety risks of Super Intelligence years before 2014 Then he started OpenAI in 2015 because he wanted a counterweight in advanced intelligence that would focus on benefiting humanity And now we’re entering the SI era....Elon was already thinking about this more than a decade ago, with a clear understanding of what it could eventually become He saw where this was going early, helped push the modern SI race forward, and spent years warning that intelligence beyond humans has to remain aligned with humanity Elon didn’t suddenly start thinking about SI because it became the biggest thing in technology He was thinking about it before everyone else knew what was coming And now we’re living through the future he was preparing for

Elon Musk explains why Tesla is being extremely careful while scaling Robotaxi The ambition is massive...but Elon’s priority is making sure they do not hurt anyone There are roughly 30,000–40,000 automotive deaths every year in the U.S., and most receive almost no national attention But if a Tesla Robotaxi injures even ONE person, it can instantly become worldwide headline news and regulators could clamp down on the entire operation So Tesla has to do two things at the same time: Scale Robotaxi as fast as humanly possible AND make sure it causes essentially no harm Elon put it very simply: “We do not want to injure anyone” “And ideally...not even run over a pet” That is the real constraint on Robotaxi right now And despite being this cautious, the rollout is accelerating fast Tesla’s authorized Cybercab fleet in Texas has nearly QUADRUPLED in just one month 45 → 169 Cybercabs And Tesla now has 589 Robotaxi vehicles authorized in Texas when you include 420 Model Ys Unsupervised Robotaxi miles have also been growing at double-digit rates week after week So Tesla is not moving slowly It’s scaling incredibly fast while trying to maintain an insanely high safety bar That’s the balance Elon is trying to solve

Falcon 9 flew more missions in ONE YEAR than the entire Space Shuttle program flew in 30 years • The entire Shuttle fleet: 135 missions from 1981–2011 • Falcon 9 in 2025 alone: 165 launches SpaceX exceeded the Shuttle program’s entire lifetime mission count by 30 launches in just 12 months And Falcon 9 also launched more times than every other orbital launch vehicle combined that year At the Shuttle program’s historical average cadence, reaching 165 missions would have taken roughly 37 years Falcon 9 did it in ONE And the hardware reuse comparison is just as insane: • Space Shuttle Discovery — 39 flights across roughly 27 years • Falcon 9 booster B1067 — 37 flights in just over 5 years B1067 first flew on June 3, 2021 By August 25, 2026, the same booster had launched and landed successfully 37 times A single Falcon 9 first stage came within TWO flights of the most-flown Shuttle orbiter’s lifetime flight count in about one-fifth of the time What once took decades of launch operations now happens within a single year of Falcon 9 cadence SpaceX turned reusability into an industrial-scale operation

Treasury Secretary Scott Bessent just made his support for Elon Musk very clear: “I am a huge Elon fan” He compared their relationship to a sports team...you can argue in the locker room, but once you step onto the field, you’re all working toward the same goal Bessent also said he’s been in regular contact with Elon and added: “He’s amazing” “He’s taking us to Mars. Look at what he’s done with SpaceX” Pretty strong words from the U.S. Treasury Secretary

Anything TSMC contributes to Terafab would basically be supplemental Terafab itself is going to be absurdly massive The goal is 1 TW of compute production every year.....around 50× the current global output of AI compute This isn’t just another fab. It’s being built for a scale the semiconductor industry simply does not operate at any known scale today

Grok Build just got another bug-fix update, this time focused heavily on permissions and tool safety Hook permission prompts now show up regardless of approval mode, WebSearch/WebFetch and image generation properly respect prompt_policy and rule decisions, and “Allow once” now really means once......only the next shell command There’s also an extra safeguard for pkill -f and pgrep -f patterns that could accidentally kill their own shell wrapper Release Notes: v1.0.49 Bug Fixes: • Hook permission prompts now appear regardless of approval mode. • Web searches, fetches, and image generation now correctly respect the prompt_policy setting and rule decisions before prompting. • Allow once in permission cards now applies only to the next shell command instead of the whole session. • pkill -f and pgrep -f commands are now rejected earlier when the pattern would kill their own shell wrapper.

Grok Build just got another update, and this one gives you much better control over complex agent runs /rewind now only shows prompts you actually typed, WebSearch and WebFetch properly respect Ask rules, and cancelling one tool call no longer kills unrelated commands running in parallel File tools are also more reliable with large, special and binary files, while authentication recovery now requires a manual resend instead of automatically retrying a refused prompt More predictable parallel execution and fewer things happening behind your back Release Notes: v1.0.48 Bug Fixes: • `/rewind` now only lists prompts you typed yourself and skips automatic runtime turns. • Custom model tables in config.toml now preserve the original model's notice banner. • `read_file`, `search_replace`, and `grep` now handle large files, special files, and binary content more reliably and surface better errors. • Ask rules for WebSearch and WebFetch now correctly show a prompt instead of auto-running. • Cancelling one tool call no longer kills commands started by other parallel calls. • Authentication recovery no longer automatically retries refused prompts after login; the sign-in block now requires manual resend.