Grok 4.5 Claims Superior Speed and Cost, While Investors Warn of AI Bubble and Infrastructure Bottlenecks

Key Takeaways

1. Elon Musk claims Grok 4.5 outperforms competitors in speed, cost, and coding, with 1.5T params being 3x cheaper than Kimi K3 (2.8T).
2. Andrew Wilkinson warns most AI companies will go to zero, advocating investments in AI infrastructure (picks-and-shovels) over frontier models.
3. @levelsio reports that Chinese models like Kimi K3 are gaining users due to fewer safety restrictions, bypassing US model limitations.
4. Multiple voices propose a "golden ticket" feature on X for users to push one tweet to all followers monthly.
5. Aravind Srinivas suggests open-source models on local hardware could disrupt the industry like Linux did to Sun Microsystems.
6. Vercel Sandbox now offers free data transfer to encourage agent deployment, while Replit's CEO shares building in-house CRM/accounting saves $30k/year.

1. AI Model Competition Heats Up

1. Elon Musk stated that Grok 4.5 (1.5T params) is significantly faster, cheaper, and better than competitors, often surpassing frontier models in practice. Kimi K3 (2.8T params) is 3x more expensive. — via 1 2 3 4 5 6
2. @tobi noted that OpenAI’s /goal feature in GPT-5.6-sol is no longer necessary as the model autonomously continues until completion. — via 1
3. @levelsio reported that Chinese model Kimi K3 easily completed a Windows XP simulation task that Claude Code refused due to safety restrictions. Users are shifting to Chinese models because US models are over-restricted. — via 1 2
4. @elonmusk also highlighted Grok TTS scoring 94 on human speech index, latency 460ms, cost $15 (7x cheaper than Eleven v3) with better quality. — via 1

2. Investment Strategy: Picks-and-Shovels Over Frontier Bets

1. Andrew Wilkinson argued that most AI companies will fail, comparing to the 1900s auto industry where only Ford and Tesla survived. He advocates investing in AI infrastructure (GPU, power, land) instead of picking model winners. — via 1
2. He highlighted that power is the real bottleneck, not GPU. Data center sites with pre-existing high-voltage lines and 24-48 month power availability are extremely scarce assets. His firm holds IREN, CRWV, NUAI, SPCX. — via 1
3. Aravind Srinivas echoed that open-source models on local hardware could disrupt the AI industry similar to how Linux and x86 disrupted Sun Microsystems. — via 1
4. @levelsio predicted that browsers will be replaced by agents in two years, with web apps becoming agent database layers. — via 1

3. Platform & Product Updates

1. Jason Fried and Clifton Sellers independently proposed X should grant each user a "golden ticket" to force one tweet to all followers monthly, ensuring important updates are seen despite algorithms. — via 1 2
2. Guillermo Rauch announced Vercel Sandbox data transfer is now free, allowing agents to clone repos, install packages, and pull artifacts without cost, aiming to boost agent publishing. — via 1
3. Amjad Masad shared that Replit helped him build in-house CRM and accounting for $100/month instead of $520, saving $30k/year, and touted Replit’s security features preventing supply-chain attacks. — via 1 2
4. Paul Graham highlighted a vaccine preventing pancreatic cancer in high-risk individuals, and AI agents achieving recursive self-improvement at Coinbase. — via 1 2