
AI Learns to Do Its Own 'Brain Surgery'
AI is on the cusp of self-improvement...
This month, MIT researchers shared the results of an experiment to see if AI agents could fix their own "brains."
AI brains are made up of "weights," or parameters that control the models' reactions to different inputs. The weights for most U.S. frontier models are proprietary. But many high-level Chinese models are "open weight." Anyone can look at their parameters – and tweak them.
The MIT researchers instructed a U.S. frontier model (OpenAI's ChatGPT-6 Astra) to read and edit an open-weight Chinese model (Qwen3.5-0.8B).
They asked Astra to improve Qwen's math skills. The results were spectacular...
With just a 0.016% change to Qwen's weights, Astra improved the model's math performance sixfold. Its instruction-following score also tripled.
In short, big AI models can read and fix the weights of other models to make them better. This may be a major change in how AI models work... and could transform humans' role in training them.
And it's another sign that the AI boom is far from over.
Smarter AI May No Longer Depend On Humans
The MIT experiment reflects a computer-science concept called "The Bitter Lesson."
In an essay published in 2019, computer scientist Richard Sutton showed that over the long run, even a general model with more computing power is better than a model based on hand-crafted human knowledge.
For example, you've likely heard of the chess-master program Deep Blue... the IBM supercomputer that beat a reigning world chess champion.
You might assume that Deep Blue was trained on classic chess strategies that it could use in any situation. But in fact, Deep Blue used a relatively simple algorithm backed by tons of computing power. It used brute force to evaluate more than 200 million chess positions a second.
In other words, IBM didn't need Deep Blue to master chess strategy. It simply had to build hardware powerful enough to run an algorithmic search.
This lesson is "bitter" because it says that humans may not have much more to teach our AI agents. All the clever tricks and lessons we gather through experience can't outdo a simple AI model with sufficient computing power.
The MIT experiment on rewriting weights is right in line with that conclusion. And it points to a big shift in AI training...
Up till now, AI agents have been trained using a system called "gradient descent." Researchers taught AI through trial and error. You'd ask models questions and tweak their weights one by one until their answers make sense.
But this process is costly and time-consuming... And it's one of the biggest obstacles to continuously improving AI.
The MIT experiment points to a new kind of AI model training: "weight-space programming." Now, a model like Astra can read all the weights in a given model and rewrite them on the fly.
This is a major step for two reasons...
AI Models Could Be Setting Up Their Own Chip Boom
First is the Jevons paradox. This economic concept states that as a process or resource gets more efficient, people use it more, not less.
Weight-space programming is a much more efficient way to build computer intelligence. So as these models become easier to train, we will likely see many more experiments and models.
Second, you might not know this, but most of the graphics processing units in AI data centers sit idle today...
These chips are available for high-demand periods – like when gradient-descent training is happening. Otherwise, they spend a lot of time as untapped capacity.
But in a "bitter lesson" world where AI models can continually monitor and tweak their own brains, those GPUs would no longer stay idle. If training becomes an automated process, AI will likely need every chip available to keep updating the brains behind its models.
Today, semiconductor stocks are under fire. The Philadelphia Semiconductor Index is more than 30% off its recent high. Take a look...
Bears are betting that chip demand has reached its peak...
But the MIT experiment points to a world where a new chip bottleneck is just around the corner. With self-improving AI eating into idle computing power, demand for semiconductors is likely to keep climbing from here.
The MIT experiment suggests that AI could soon build itself far better than any human could. And that can power a bull run in chips much higher than anyone expects.
Good investing,
Sean Michael Cummings
Further Reading
Bank of America has thousands of employees working to protect its software from AI. Smaller businesses don't have that luxury. As increasingly powerful AI tools continue to uncover system vulnerabilities, companies will need outside help to protect themselves. That's creating a new source of demand for cybersecurity businesses.
The race to build humanoid robots is about more than making machines that can walk, talk, and perform physical tasks. It depends on real-world data. The companies that figure out how to collect enough of it could pull ahead in the technology market of the future.
Market Notes
HIGHS AND LOWS
NEW HIGHS OF NOTE LAST WEEK
Tradeweb Markets (TW)... online trading platform
Johnson & Johnson (JNJ)... health titan
Becton Dickinson (BDX)... needles and syringes
Edwards Lifesciences (EW)... medical devices
Apple (AAPL)... Big Tech
Zeta Global (ZETA)... AI-powered marketing
Columbia Sportswear (COLM)... outdoor apparel
Costco Wholesale (COST)... membership-only stores
Crocs (CROX)... footwear
Coca-Cola (KO)... soft drinks
Procter & Gamble (PG)... consumer goods
Republic Services (RSG)... waste disposal
Waste Connections (WCN)... trash and recycling
NEW LOWS OF NOTE LAST WEEK
Annaly Capital Management (NLY)... "virtual bank"
Coinbase Global (COIN)... cryptocurrency exchange
Crispr Therapeutics (CRSP)... gene editing
Flowers Foods (FLO)... baked goods
Corteva (CTVA)... agriculture
Newmark (NMRK)... real estate
