An engineer with no machine-learning background handed OpenAI's newly released GPT-5.6 Sol the reins to run its own experiments, and the model produced a tiny local autocorrect system that outperforms Sol itself—at a total cash cost of $0.
July 2026 · GPT-5.6 Sol experiment
An AI built an autocorrect that beat itself — for $0
An engineer with no machine-learning background handed GPT-5.6 Sol the wheel. The model planned, coded, and fine-tuned a tiny on-device autocorrect that edged out Sol itself — at no cash cost beyond a single API quota reset.
$0
Total cash cost — nothing beyond one API quota reset
1.7B
Parameter model, a T5Gemma encoder-decoder fine-tuned locally on a MacBook
91.02%
Error-reduction rate — the new local model's headline score
Error-reduction rate, side by side
Higher is better — the tiny local model tops even the flagship
91.02%
LOCAL
New local
How the model drove the whole pipeline
1 · Shortlist
Read the literature and picked candidate base architectures
→
2 · Simulate
Built a physical keyboard simulator to generate realistic typo data
→
3 · Fine-tune
Custom loss + beam search, trained locally on a MacBook via MLX
→
4 · Iterate
Debugged tokenizer & loss bias on its own; validated on unseen words
Why it matters
A non-expert orchestrated a full autonomous research loop — planning, coding, refining — at near-zero cost, pointing to accessible, private, fully local model workflows.
The caveats
It's one narrow task, and the edge over Sol is slim (+0.46 pts). The demo speaks to accessibility — not a decisive leap in raw model quality.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…