
𝗜 𝗧𝗿𝗶𝗲𝗱 𝗪𝗲𝗶𝗴𝗵𝘁-𝗦𝗽𝗮𝗰𝗲 𝗣𝗿𝗼𝗴𝗿𝗮𝗺𝗺𝗶𝗻𝗴 𝗼𝗻 𝗠𝘆 𝗗𝗚𝗫 𝗦𝗽𝗮𝗿𝗸 𝗪𝗶𝘁𝗵 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲 𝟱.𝟮 𝗮𝘀 𝘁𝗵𝗲 𝗪𝗿𝗶𝘁𝗲𝗿, 𝗮𝗻𝗱 𝗔𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝗪𝗲𝗻𝘁 𝗙𝗿𝗼𝗺 𝟱𝟯% 𝘁𝗼 𝟳𝟭% Spent the weekend on weight-space programming. The idea: one LLM reads another network's weights and edits them directly. No gradients, no fine-tuning. I used dspy-wsp from cmpnd-ai, which builds on "Sparks of Weight-Space Programming" by Takehi, Teoh, Zhang and Khattab. The Two Models: The model being fixed is LiquidAI d1-omni-600M, a small open model (about 587M weights) running on the Spark's GPU. It reads a business record, say an invoice or a security alert, and answers 5 questions about it. The model doing the fixing is called the writer. It gets a locked Python sandbox with the small model inside and 25 runs over the data. 𝗙𝗶𝗿𝘀𝘁 𝗧𝗿𝘆: 𝗠𝘆 𝗟𝗼𝗰𝗮𝗹 𝗤𝘄𝗲𝗻𝟯.𝟴-𝟮𝟳𝗕: I tried my local Qwen3.8-27B first because it's free. Went to bed and let it work. 𝟲.𝟵 𝗵𝗼𝘂𝗿𝘀 𝗹𝗮𝘁𝗲𝗿 𝗶𝘁 𝗵𝗮𝗱𝗻'𝘁 𝘀𝗮𝘃𝗲𝗱 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗲𝗱𝗶𝘁. +0.0. (To be fair, I gave it 90 code runs instead of the default 300, and 36 of them crashed.) 𝗦𝗲𝗰𝗼𝗻𝗱 𝗧𝗿𝘆: 𝗖𝗹𝗮𝘂𝗱𝗲 𝗙𝗮𝗯𝗹𝗲 𝟱.𝟮 𝗧𝗵𝗿𝗼𝘂𝗴𝗵 𝗢𝗽𝗲𝗻𝗥𝗼𝘂𝘁𝗲𝗿: Then I made Claude Fable 5.2 the writer, through OpenRouter. It took 𝟴𝟴 𝗺𝗶𝗻𝘂𝘁𝗲𝘀. It started by testing the small model and found it right about half the time. Then it spent 95 Python runs looking inside. What it found: the facts needed for the right answers were already in the model. The last layer, the scorer (it turns what the model read into a score for each option), just wasn't using them. The original model called all 100 test invoices "Routine", overdue or not. So it 𝗱𝗶𝗱𝗻'𝘁 𝘁𝗲𝗮𝗰𝗵 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗮𝗻𝘆𝘁𝗵𝗶𝗻𝗴 𝗻𝗲𝘄. It rewired 144 of the scorer's 1,024 units, 𝟭𝟰𝟳,𝟳𝟰𝟰 𝘄𝗲𝗶𝗴𝗵𝘁𝘀, 𝗮𝗯𝗼𝘂𝘁 𝟭 𝗶𝗻 𝟰,𝟬𝟬𝟬, using one direct calculation on 900 practice cases. It checked the result on 300 cases it hadn't used, improved it twice, kept the best version and saved it as a small patch. The original model is untouched. On 400 test cases it never saw (2,000 answers), accuracy went from 𝟱𝟯.𝟯𝟱% 𝘁𝗼 𝟳𝟬.𝟵𝟱% (+17.6, ±1.2). 257 cases got better and 𝟱𝟭 𝗴𝗼𝘁 𝘄𝗼𝗿𝘀𝗲. The dspy-wsp team's own run with gpt-6-astra got 68.3% on the same test. The picture I keep coming back to is a student who read the whole book, then ticks the same box on every question. More lessons won't help. You fix the answer sheet. That also tells you the ceiling. Fable estimated this kind of last-layer fix tops out around 74 to 75%, because two cues never reach that layer at all. Cost: The whole thing cost $𝟭𝟳: about $14 for the main run and $3 for test runs. Prompt caching did a lot of the work. 95% of the input came from cache, which costs 1/40 of the normal price, so without it I'd guess around $50 for the main run. One gotcha: DSPy's default cache setting sent nothing through OpenRouter. It only kicked in once each call marked its newest message. I would not call it a clean win: I wouldn't call it a clean win over astra, though. Each writer ran once and astra's number is the team's published run, so Fable is 𝘀𝗹𝗶𝗴𝗵𝘁𝗹𝘆 𝗮𝗵𝗲𝗮𝗱 𝗮𝘁 𝗯𝗲𝘀𝘁. Astra was cheaper too, at $7.63. And 70.95% counts the option the model rates most likely. The answer the program actually returns only reaches 𝟳𝟬.𝟳% after DSPy's ReAnchor refits its cut points. Without that it's 𝟲𝟬.𝟮%. What stuck with me is the gap between writers. A local 27B model got nothing out of the same job that Fable turned into +17.6. The README points the same way: on this model gpt-6.1-sol gained 5.8 points and gpt-6-astra 16.1 (on returned answers). I think the writer is most of the game. Claude Code did the setup and runs, and built the live demo of both models side by side you see in the video. I'll put the code, demo and notes in a repo. 𝘚𝘢𝘮𝘦 𝘵𝘰𝘰𝘭, 𝘴𝘢𝘮𝘦 𝘮𝘰𝘥𝘦𝘭, 𝘴𝘢𝘮𝘦 𝘥𝘢𝘵𝘢. 𝘖𝘯𝘭𝘺 𝘵𝘩𝘦 𝘸𝘳𝘪𝘵𝘦𝘳 𝘤𝘩𝘢𝘯𝘨𝘦𝘥, 𝘢𝘯𝘥 𝘪𝘵 𝘸𝘦𝘯𝘵 𝘧𝘳𝘰𝘮 +0.0 𝘵𝘰 +17.6. Links in reply 👇




