AI Video Realism and Agent Infrastructure Surge as OpenAI Projects Massive Token Growth

Key Takeaways

• ByteDance's Seedance 2.5 can generate near-identical video likeness from a few reference images, with a 15-second clip taking about 4 minutes; @levelsio's fake 'iPhone 18 Pro Fold' promo demonstrates the realism. — via 1 2
• Sam Altman marks GPT-4's four-year training anniversary and projects token consumption could grow from ~100k average per person today to 500 billion in six years. — via 1 2
• Agent infrastructure is maturing: Vercel adds bill protection, Buzz fixes managed-agent I/O, and Overeasy layers a recoverable filesystem onto S3 logs. — via 1 2 3
• Founders are preparing for agent-heavy organizations where context, not humans, is the durable asset. — via 1
• @levelsio warns AI-generated content could pollute website reviews, making authenticity verification a core problem. — via 1

1. AI Models, Video Realism, and Token Growth

1. ByteDance's Seedance 2.5 produces video from a few reference images at near-photo-level likeness, with a 15-second clip taking about 4 minutes. It is live on Photo AI and available through Vercel AI Gateway, where Guillermo Rauch described the result as 'Dreamcore.' — via 1 2
2. @levelsio showed an AI-generated promo for a fake 'iPhone 18 Pro Fold,' created from voice and a few photos, claiming it is almost indistinguishable from a real ad. This illustrates both the creative upside and the authenticity risk of AI video. — via 1
3. Grok Imagine Image 2.0 is now ranked #2 on the leaderboard and is in preview through Vercel AI Gateway. Rauch called it an excellent model, showing how fast image generation is commoditizing. — via 1
4. Sam Altman noted GPT-4 completed training four years ago today, a reminder of how quickly the current AI wave has moved from model milestones to product deployment. It also signals OpenAI's willingness to celebrate durability rather than just launch-day buzz. — via 1
5. Garry Tan amplified Sam Altman's projection that average per-person token usage has jumped from ~0 to ~100,000 in six years. If the curve holds, average usage could reach 500 billion and top users could hit quadrillion-level usage, pushing companies to own their data, models, and compute. — via 1 2

2. Agent Infrastructure and the Future of Work

1. Greg Isenberg argues the future org chart is a thin layer of humans for strategy, creativity, and judgment, with a broad base of agents handling customer support, sales, research, marketing, and operations. Managing agents means managing context, and that is the skill millions will need now. — via 1
2. Tobi Lutke points to Overeasy, a persistent filesystem overlay on S3 immutable logs that supports forking, cross-machine recovery, and rollback to any historical state. He frames all mutable state as eventually reducible to state = memo { f(log) }. — via 1
3. Yohei Nakajima proposes treating a repository as persistent project state: models, agents, tools, and sessions can change while new workers inspect prior work, take limited roles, and leave verifiable handoffs. — via 1
4. Buzz Desktop v0.5.8 fixes message read/write issues with managed agents like Claude and Codex, improving long-session handoffs. This addresses real friction in running agentic coding workflows. — via 1
5. Vercel now prevents surprise cloud bills with soft/hard limits, anomaly alerts, function recursion protection, queryable billing APIs, and DDoS mitigation included in all plans. These controls target the biggest concern for developers moving autonomous agents to production. — via 1
6. Amjad Masad made an unverified claim that a rogue OpenAI agent independently developed Kantian ethics, while also pointing to 'what comes after prompt-based AI tools' as the next direction. — via 1 2

3. Founder Insights and Startup Signals

1. Paul Graham told a startup that post-launch growth peaks are not meaningful and asked for the pre-peak rate; the founder answered 40% week-over-week. That pre-peak number is the signal experienced founders look for. — via 1
2. Paul Graham also flagged Firecrawl entering GitHub's all-time Top 50, connecting it to the need for AI agents to get better web context. — via 1
3. Garry Tan argues many of America's biggest tech companies were born from fixing San Francisco's broken systems, with housing creating Airbnb, transit creating Uber, and social life creating Hinge and Instagram. His call now is to use AI to make SF 'cool' again. — via 1
4. Naval argues wealth gaps are smaller than people think because of mass production, but status is what people actually want, and it is a zero-sum, destructive game. — via 1
5. Sahil Lavingia called '996' a pre-agent phenomenon, suggesting extreme human office hours may become less relevant as agentic work scales. — via 1
6. Alex Hormozi's wealth formula: hold a higher internal standard than your clients and have the courage to charge high fees. His framing ties pricing power directly to internal standards. — via 1

4. Media, Business, and Online Services

1. Athlos London has signed a BBC partnership to broadcast the all-women's league's UK debut on September 18 from StoneX Stadium. It gives the league a national broadcast platform ahead of its first event. — via 1
2. The IRS online business tax account now lets organizations view digital notices, make payments, and manage federal tax information online. This is part of the broader movement to modernize government self-service tools. — via 1
3. @levelsio argues AI agents can't reliably judge new products from websites because AI-generated text and video pollute reviews, making 'is it real?' the core question. He also predicts LLM apps may go free in exchange for user data, with models actively asking users questions to gather training data; if people stop creating fresh content, AI companies may shift to chat logs and real-world experiences. — via 1 2 3