DeepSeek R1 Code Optimization and a $3 Visual Counting Experiment

I see a recent DeepSeek R1 coding example as a sign of self-improving AI. In my view, systems able to contribute to AI research could eventually drive recursive improvement and an intelligence explosion. That is a prediction about the future; the example I show involved a developer directing and testing a specific code change.

I think models such as o1, o3 and DeepSeek R1 are approaching a level where they can contribute to new knowledge. I also want to show a separate team’s experiment that elicited a reasoning-like “aha moment” on a narrow task for less than $3, following a separate $30 example I discussed earlier.

Simon Willison’s blog highlights a llama.cpp pull request by Xuan-Son Nguyen reporting a roughly twofold speed improvement for particular WebAssembly dot-product functions using SIMD. Nguyen said DeepSeek R1 wrote 99% of the code, while he developed tests and wrote prompts through trial and error. This is a change to inference software, not a demonstration that the model autonomously changed its own weights.

Willison also describes using R1 to rewrite an llm_groq.py plugin to follow a cached-model JSON pattern in llm_mistral.py. He says he tried the same task with o1 and thought R1 did better. It is an individual programming experience rather than a broad benchmark comparison.

The developer used the free chat.deepseek.com interface, with responses reportedly taking around three to five minutes. The iterative prompt described the problem and included the last failed attempt, asking the model to improve it. One example asks for a conversion from C++ ARM NEON SIMD to WebAssembly SIMD, with existing code supplied as context.

The prompts let the model look for code optimizations, and another example shows it working through how to apply a pattern from the Mistral plugin to Groq. I find this striking and imagine many agents continually searching for improvements. That imagined autonomous process goes beyond the human-guided examples shown here.

I return to an intelligence-explosion chart and argue that automated AI research could put us close to superintelligence. I am conflicted about whether progress would be a sudden takeoff or more gradual. Yann LeCun, for example, argues that AGI will emerge progressively and be reproduced by others relatively quickly once it appears.

I think open-source releases can accelerate the spread of advances, and I regard DeepSeek R1 as an important example. I also cite Sam Altman saying in a Reddit AMA that R1 is a very good model and that OpenAI would produce better models but maintain a smaller lead than in previous years.

The second example is R1-V, introduced by Liang Chen. The team uses reinforcement learning with verifiable rewards to encourage vision-language models to learn counting. I compare it with the earlier $30 narrow experiment and DeepSeek R1’s approach, while the specific task and setup here are different.

A verifiable reward gives training a way to check an answer: arithmetic has a definite answer, whereas a favorite color is a matter of opinion. I argue this makes the technique especially appealing for well-defined problems in science, technology, engineering and mathematics. The example is about tasks with checkable rewards, not a claim that reinforcement learning cannot be used elsewhere.

Chen reports that a 2-billion-parameter model surpassed a 72-billion-parameter model on the counting task after 100 training steps costing less than $3. This is a narrow comparison, not evidence that the smaller model is generally better. I speculate that many small, specialized models could someday be selected by a router in response to a user’s request.

The displayed out-of-distribution evaluation on CLEVR CoGenT-B counting reports 53% for the original Qwen2-VL 2B model and 99% after R1-V training, compared with 94% for Qwen2-VL 72B. The chart also shows 30% for a supervised fine-tuning variant. These figures describe the shown evaluation and its particular training setup.

The R1-V team describes a goal of building a general framework for reinforcement learning with verifiable rewards in vision-language models. I imagine applying such tools to other tasks where rewards can be checked, and see the community’s ability to reproduce and extend ideas as a strength of open source.