Meta's Muse Wonder Debuts as Grok Bot, Perplexity Projects and Agent Tooling Accelerate

Key Takeaways

1. Meta released Muse Wonder, a new frontier model positioned for state-of-the-art personal and professional agentic work, with rapid third-party validation and immediate availability on Vercel AI Gateway.
2. Grok Bot drew heavy user-reported usage across email, travel booking, code review and autonomous publishing, while a new memory-transfer feature lets users migrate context from Claude or ChatGPT with item-by-item review.
3. Perplexity Projects was described as a Git-forge-style file system for agents, with sandboxed work, shared repositories and a reported sub-2% commit conflict rate.
4. Google DeepMind researchers used AI to solve nine previously open Erdős problems and prove 44 integer-sequence conjectures, with Lean-based formal verification central to the workflow.
5. SpaceX-related posts covered Starlink deployments, a Falcon 9 Dragon resupply launch and a user-compiled Starship flight retrospective, alongside Musk's manufacturing-scale commentary.
6. Startup and operator commentary focused on profitability over fundraising, AI adoption by offline businesses, and the shifting bottleneck from building to distribution.

1. AI Models and Agent Infrastructure

1. Meta released Muse Wonder, described as a new frontier model series targeting state-of-the-art performance in personal and professional agentic work. It was made available on Vercel AI Gateway via fx --model meta/muse-wonder-1.0 and positioned for frontier coding, long-running agents and advanced computer use. Multiple forwarded assessments said it is very strong, hard to stump, and that "Meta is back," with one noting it found a real security vulnerability other agents missed and excels at design and visualization. Another forwarded post said Muse Wonder is the result of more than a year of MSL research and infrastructure work, scaling pretraining and RL since the April release of muse spark. — via 1 2 3 4 5 6 7
2. Grok Bot drew a wave of user-reported use cases: replacing search, Hermes automation and Cursor cloud agents; reviewing PRs and merging code; connecting to machines via Tailscale and SSH; launching Claude Code or Codex agents in the terminal; planning and booking a Lisbon trip with card payment; handling roughly 20 emails, calendar, proposals, lunch and hotel booking and airport taxi in a 30-minute "call"; fixing a customer bug and publishing autonomously; and running 62 agents with 29 scheduled routines. Musk amplified the feedback with short endorsements including "Grok @Bot rocks" and "It gets better every day." A separate forwarded feature description said Grok Bot memory transfer can migrate what Claude or ChatGPT knows about a user by pasting a prompt into the old chat or importing a full data export, with every item reviewed jointly and only kept entries saved with source and timestamp. — via 1 2 3 4 5 6 7 8 9 10 11
3. Perplexity Projects was described as building its file system as a Git forge for agents: each agent works in an isolated sandbox and commits to a shared repository so humans and other agents can continue building across sessions. The author said each project is a Git repository, avoiding GitHub overhead because no human approval is required, with short-lived sandboxes committing to a shared repository on S3 and a commit conflict rate below 2%. Agents reportedly use plain ls and grep, beating semantic search 63% to 37% on answer accuracy in financial document analysis. — via 1 2
4. A Google DeepMind team used AI to solve nine previously unsolved Erdős problems, some open for 56 years, and proved 44 previously unproven integer-sequence conjectures. The results are described as real mathematics rather than benchmarks. The agent generated proofs in Lean, with every step passing a formal checker and the model receiving feedback to retry on failure. A simplified approach — an LLM repeatedly generating proofs checked by Lean — also solved all nine Erdős problems, though the original complex system performed better on some especially hard problems and the comparison came after the original discovery. The author suggested progress in autonomous mathematics may come more from connecting language models to reliable verification tools than from increasing agent complexity, while noting other AI-for-Science fields still need equally reliable verification. — via 1

2. SpaceX, Starlink and Manufacturing Scale

1. Musk forwarded a claim that once China masters a product's manufacturing method, its scaling capacity often exceeds the rest of the world combined, citing lithium batteries and solar as examples. SpaceX was called a rare counterexample, with orbital mass about an order of magnitude above the rest of the world, while China currently accounts for most orbital mass outside SpaceX. The same post said RAM was a low-margin commodity about three years ago, that LP5 and LP6 were designed to a common standard, and that China appears to be solving this quickly, while newer HBM architectures, SRAM layers under logic and 3D-stacked RAM with logic cooling will be harder. — via 1
2. Starlink-related posts included the Gyeongsangbuk-do government in South Korea introducing Starlink into fire services for remote mountain areas, with the fire headquarters equipping a command vehicle for a pilot to share footage and coordinate during wildfires, floods and typhoons. Another post said Starlink helped rescue five people at sea after the boat Magic Bus sank far off the North Carolina coast; the captain used a Starlink Mini connected to a phone to call the U.S. Coast Guard after radio calls failed, and a rescue helicopter brought all five back safely. A separate post said the Lagos State government distributed 1,500 Starlink kits to orphanages and children's homes, with BEB collaborating with the state government. — via 1 2 3
3. A forwarded post said a Falcon 9 launched Dragon to the space station, with launch scheduled for 6:33 a.m. Eastern that day, carrying more than 6,300 pounds of science, supplies and roll-out solar arrays for the Expedition 75 crew. Musk also forwarded a Falcon 9 launch of 27 Starlink satellites from California and confirmed deployment of all 27. — via 1 2 3 4 5
4. A forwarded long post described using @bot with only official SpaceX material to compile raw Starship launch footage from Flight 1 through Flight 14 into an 11-minute video. The author recapped each flight, including Flight 1 breaking up about 24 miles up and damaging the pad on April 20, 2023; Flight 2 lighting all 33 booster engines and first hot-staging on November 18, 2023; Flight 3's first reentry attempt and payload door opening on March 14, 2024; Flight 4's controlled booster splashdown and ship return with damaged flaps on June 6, 2024; Flight 5's first mid-air catch of Super Heavy on October 13, 2024; Flight 6's first Raptor relight in space on November 19, 2024; Flight 7's new-generation Starship first flight and second booster catch but ship breakup on January 16, 2025; Flight 8's third booster catch but ship lost again on March 6, 2025; Flight 9's first launch of a previously flown Super Heavy on May 27, 2025; Flight 10's first deployment of eight Starlink satellite simulators on August 26, 2025; Flight 11 as the last flight of that version on October 13, 2025; Flight 12 as Starship V3's first flight on May 22, 2026; Flight 13's first deployment of 20 real next-generation Starlink V3 satellites on July 24, 2026; and Flight 14 reaching orbit and deploying 26 real Starlink V3 satellites on September 28, 2026, with Musk later confirming all 26 were operating. The author also said Musk said Friday that a SpaceX market cap above $1 quadrillion is a possible outcome, after previously not believing it, and that $SPCX is currently worth about $2.2 trillion. — via 1

3. Startups, Profitability and AI Adoption

1. Garry Tan argued founders should focus on profitability and run companies with maximum agent loops rather than relying on equity fundraising: $10 million in annual profit reinvested is equivalent to an undiluted annual Series A, and the best startups already do this. A forwarded post said A and B rounds are the hardest for founders now, with vertical AI SaaS companies reaching $1M–$3M ARR in two years or $5M–$10M ARR in three to four years, and that unfinished A/B rounds do not naturally happen after a year of slight progress. Another forwarded post agreed founders should stop worrying about capital markets and focus on building as profitable a company as possible. — via 1 2
2. Jason Fried said if he were not doing his current work, he would help offline businesses transition to AI and automate, echoing a post that said helping offline businesses adopt AI can be very lucrative because once one competitor adopts AI, others are forced to follow. Fried also said he would love for 37signals to return to customer service because improving existing things has never been more fun, while he has less interest in selling, servicing, marketing and maintaining entirely new products. — via 1 2
3. Clifton Sellers said large retailers cut 20% of SKUs this year while expecting sales growth, and advised service providers with five packages to cut three and watch conversion. He argued the reason "every founder will become a marketer" keeps trending is that building costs are falling and the bottleneck has shifted to whether you can get 100 strangers interested. — via 1 2
4. Aravind Srinivas said a single-person company spends more than $1 million per year on Perplexity, with the entrepreneur earning $3 million per year by automating newsletters for others. He also said Perplexity's Computer classified urban land use from aerial photos in Phoenix, analyzing 203,883 USDA public imagery tiles with the Decider API and reaching 98.8% accuracy against OpenStreetMap labels at an API cost of $9.63. — via 1 2
5. Yohei Nakajima highlighted the first autonomous AI research with real human participants, congratulating ExScienceAI on a peer-reviewed paper on the cover of Wiley's Advanced Science. The run lasted about 17 hours, involved 50+ agents, consumed 32.5 million tokens, wrote about 7,696 lines of code, and cost about $114 in marginal compute (±30) plus $4,500 in human costs across 288 participants. The paper is titled "Closing the Empirical Loop: Autonomous AI Agents Conduct End-to-end Research With Human Participants." — via 1