Elon Musk calls GrokBot's serving costs unsustainable and reignites the subsidy-era debate. Meanwhile, an MIT paper shows a frontier model editing a smaller model's weights directly, and Anthropic's updated S-1 says it may imminently hit its RSI danger threshold. The three-letter acronym is everywhere. The harder question is what it means when it arrives.
We're so busy asking whether RSI is here that we're not asking what it means.
Sundar Pichai expects RSI to drive acceleration over the next 12 months. MIT researchers show GPT-6 Astra rewriting a smaller model's weights. Anthropic's S-1 warns it may imminently cross its RSI danger threshold, and Wall Street is reading the charts. Yet better models aren't a straight line to revenue: enterprises still consume a vanishingly small share of available intelligence because adoption is hard. Like AGI before it, RSI risks becoming a definitional sideshow. The more useful question is what the world looks like once it arrives.
— Sundar Pichai, Alphabet CEO, on the company's earnings call. On Alphabet's earnings call, Pichai said AI is already helping Google increase its training and launch velocity. Recursive self-improvement is no longer a fringe idea. It is showing up in CEO earnings commentary.
After GrokBot users hit usage limits they had never run into before, Elon Musk replied that the cost of serving the product was unsustainably high relative to revenue. He gave no further details, but for market bears the post landed with the force of a bomb.
Weeks earlier, Musk had said he'd never seen anything grow as fast as GrokBot and that the lack of a free tier proved people derived real value. Days earlier, he had posted that saving money is pointless given superintelligent abundance. Critics struggled to square either message with the new cost warning.
Many users framed it as shrinkflation: they kept paying for the same plan and got less with zero notice, then were told they had picked the wrong tier. The lesson for any AI vendor tightening limits is to announce changes ahead of time so customers can adjust.
Observers called it a win for Dots and Duet. The makers of Hermes jumped in with step-by-step instructions for migrating from GrokBot to their agent.
Dax from OpenCode argued people don't want to pay for personal agents, so they have to be top-of-funnel for something else that nobody has identified yet. This echoes the fall 2025 bubble debate over ChatGPT's costly free users, and skeptics quickly extended it to Meta, whose position in the AI race rests heavily on Muse.
Skeptics called it the moment the music stops: VC money runs out, prices rise, users drop off, valuations plummet. One newsletter cited an analysis estimating that a full month on a $200 ChatGPT Pro plan would cost about $14,000 at regular API prices.
NLW is no AI bear, but he welcomes the skeptics. It would be far more dangerous if everyone were too enamored of Anthropic and OpenAI to ask critical questions. He also notes that bearish analysis varies widely in quality, and permabears who keep moving the goalposts may not be the best guide.
Some users find more efficient models good enough and stay on $20–30 plans. Others get enough value to pay for the intelligence they consume as a cost of doing business. For NLW, API spend is net new money, still far below the cost of equivalent human labor and the time to recruit and manage it.
Will users abandon AI once they pay API rates, or is revenue already too concentrated among power users who pay via the API? The two claims sit awkwardly together. The more interesting middle ground, argued by Chamath Palihapitiya, is that too many power users are still subsidized by VC dollars.
He posted "Yikes" after CoreWeave's bonds had a bad day, and later claimed debt markets had almost entirely shut off for companies without substantial non-AI-data-center revenue. Some suspect 4D chess aimed at cheap acquisitions or hurting rivals, but no one knows.
a16z's Olivia Moore points to Meta's 3.64 billion daily users and more than 500 million SMBs already marketing on its platforms. Meta plans to connect Muse to ad analytics plus tools like QuickBooks and Stripe, which could make Muse a nervous system for SMBs.
SemiAnalysis's new GDPX benchmark covers tasks like chip design, finance and biology. Opus 5.5 scored highest at 12% for $51 per task, while Gemini 4 Carbon hit 10.9% for $16, roughly 4× cheaper. The labs are now focused on using efficient models to cut personal-agent costs, as they did with coding agents.
OpenAI's LifeVal measures tasks like finding childcare, budgeting and ordering groceries. Scores rose from 49.3% for GPT-5.2 to 80.6% for GPT-6 Astra. As with coding, once a capability becomes a measured RL target it tends to improve fast and get cheaper, and some read the benchmark as a prelude to a public personal-agent launch.
MIT researchers found GPT-6 Astra could read and adjust the weights of Qwen 3.5 0.8B. By changing just 0.016% of them, it improved math benchmark scores 6× and instruction following 3×. The researchers say frontier LLMs are surprisingly good at natively reading and writing neural network weights.
Asked to make Qwen better at math, Astra spotted the model's known reasoning-loop problem, located the cause within a budget of 50 forward passes, and edited it out. That boosted other capabilities too. Astra also wrote small game-playing networks from scratch that outplayed networks trained with PPO for 100 million steps.
A senior Google researcher quickly reproduced the results and called it a step-change capability that appeared only in the latest models. The edited models are tiny, though: 20,486 parameters in his experiment, versus more than a trillion for frontier models. He frames it as 2010s-style interpretability work now done by models at a scale beyond humans.
The main critique is that the 6× math gain came from fixing a loop bug, not from improving general capability. Others say the approach is undersold, reporting that Claude models can write functioning neural net weights from scratch. RSI believers argue interpretability is dual-use: models good at optimization are good at optimizing themselves.
New language in Anthropic's updated S-1 says internal estimates indicate it may soon hit its threshold of dramatically accelerated progress attributable to AI: as much progress in one year as two years at the 2018–2024 pace. The company is considering responsible scaling policy updates aimed at keeping humans in control of frontier development.
— Anthropic, updated S-1 filing. The full line adds "without meaningful human supervision or understanding." It is a frontier lab, in an IPO filing, describing the limits it plans to put on AI building its own successors.
The Wall Street Journal ran Anthropic's chart showing AI now leading 26% of research, echoing Jack Clark's view that RSI might be the most important technical trend in the world. The data is from August and was boosted by David Sacks, who wondered where the number stands now. With the AI narrative souring, investors are watching RSI as a source of the next leg of growth.
The unspoken assumption is that improved models drive adoption. That held for true step changes like the Opus 4.5 jump that kicked off the agentic coding era. Without an entirely new capability coming online, NLW doesn't think the evidence says capability alone drives adoption.
Whether these business models work comes down to how much intelligence enterprises consume, and today that is a vanishingly small share of what they could. The reason is inertia, lumbering corporate systems and hard adoption, not insufficient models. The bear reading is that we aren't using it; the bull reading is that we aren't using it yet.
The debate is stuck on whether RSI is here yet. The only group offering a vision of the world after RSI is the AI safety camp that wants it banned. Like AGI before it, the acronym may be absorbing attention that should go to understanding what it will mean when it arrives.
