MiMo code
Xiaomi’s MiMo Code supposedly beats Claude Code at ultra-long, 200+ step tasks. Give it a try.
If you pay, you can check account balances on the Xiaomi site.
QWEN 3.6 27B Multi-Token Prediction model
MTP (multi-token prediction) is an AI training paradigm where large language models (LLMs) are trained to predict several future tokens concurrently at each position, rather than predicting only the single next token.
llama.cpp has added the feature and some report almost doubling their tokens per second. I found turning on MTP (multi-token prediction) did indeed improve performance and tokens per second.
You’ll definitely need to pull the latest, greatest llama.cpp code and re-compile it as the mtp support is very new. You also need to download a version of the Qwen 3.6 27B model with MTP support here. My current command line for a 5090 is not fully optimized, but here it is:
set LLAMA_CACHE=unsloth\Qwen3.6-27B-MTP-GGUF
llama-server.exe -hf unsloth/Qwen3.6-27B-MTP-GGUF:UD-Q4_K_XL -ngl 99 -c 200000 -fa on -np 1 --spec-type draft-mtp --spec-draft-n-max 3 -ctk q8_0 -ctv q8_0 -b 1024 -ub 256 -t 16 --no-mmap --temp 0.6 --top-p 0.95 --top-k 20 --presence_penalty 0.0 --repeat-penalty 1.0 --reasoning on --host 0.0.0.0 --port 8001 --metrics --slots --props --chat-template-kwargs "{\"preserve_thinking\":true}"I’ve tried this with both Q4 and Q8, and it seems to definitely speed things up.
Quantum Computer hooked to AI
“The results reported here constitute, to our knowledge, the first demonstration of end-to-end quantum enhancement of a production-scale, widely-deployed LLM on real superconducting quantum hardware for autoregressive language generation,” the scientists wrote in the study. “Their significance lies not in the magnitude of the perplexity improvements — which will grow with hardware fidelity and qubit count — but in the fact that they exist at all.”
IBM researchers have demonstrated a new way of using a quantum computer to fine-tune a pretrained large language models (LLM) and have achieved measurable improvements in the model’s ability to forecast text.
It did this by a quantum-assisted optimization technique that reduced the model’s “perplexity” (the metric for how well an AI predicts the next word). The model became more accurate at answering questions — including some it previously got wrong.
This is an amazing development not only for the possibility of quantum-enhanced LLM’s – but is probably the first real demonstration of quantum computing demonstrating a clear advantage of classical computing. Quantum computing has struggled to demonstrate its advantages in real-world applications. This appears to be the first real-world application that showed clear improvements.
It makes me wonder if this works so well because LLMs and AI rely on the fuzzy statistical domains that quantum computing’s big issue of stability matches well with. Perhaps it could even flip things on it’s head. Perhaps an specially trained AI system might turn out to be the perfect API (if you will) for determining outcomes from fuzzy quantum computations.
Link:
Google Stitch for all your UI needs
Is Figma dead?
Google Stitch is a prompt driven complete AI designer that allows you to develop full UI designs almost instantly. But it’s not some isolated tool. You can export your UI designs from the system into Claude Code, Cursor, and any other coding agent.
Does it work? Yes. I re-created most of the UI of an app I wrote in less than 15 minutes. It was so easy, I literally gasped when it generated almost a better version of my own app.
Have inspiration? Point Stitch at any website and it will create it’s own version for you as a starting point.
Then you can go in and change the color palette and automatically create light/dark modes.You can then use the editor to add and change individual elements: change images, logos, create animations, add and remove App Store, Web, Marketing Kit, and Accessibility assets.
With the Direct Edit pencil you can click any text or swap any images for fine control – all without charging credits.
If Figma and UI coders aren’t dead, they definitely need to up their game to use these kinds of tools or they soon will be.
AI Python/pip wheels
What a treasure when I stumbled upon this! I was pulling my hair out trying to get a working copy of Flash Attention compiled for Qwen-TTS agent I was trying to run locally on Windows.
Compiling some AI packages for Windows systems can be notoriously horrible experiences (looking at you flash-attn). Here you can tell it which version of CUDA, Pytorch, and Python and then find a pre-compiled wheel for that configuration.
Get the versions of your tools by using these command lines:
- > python –version
- > python -c “import torch; print(torch.version)”
- > nvcc –version
Use AI to teach you better AI prompting techniques
It’s becoming more and more common to use AI to help you understand what AI is doing. Task decomposition has been around a long time with asking AI to break things into steps to implement. This is kind of a new technique I had not seen: prompt decomposition. The idea is to use AI to help generate a good prompt that will get you want you want.
This could be used on cheaper local models you run for free to generate the prompts you use on expensive models.
The query goes something like this:
I want to create a high-quality prompt for this task:
[TASK]
Before writing the prompt, identify the 5–7 high-leverage prompt dimensions for this task — the core variables, constraints, context, output requirements, or stylistic choices that will most determine the quality of the result.
For each dimension, briefly explain:
1. Why it matters
2. What tradeoff or decision it controls
3. How it should influence the final prompt
Then turn those dimensions into a polished, copy-ready prompt.
The final prompt should be clear, specific, and structured. It should include the necessary context, role, task instructions, constraints, output format, quality criteria, and guidance for handling ambiguity.
Automated sand delivery
Driverless autonomous trucks are becoming a thing in remote areas of Texas used to feed an endless sand supply to fracking wells. AI and fracking. How could that go wrong? 😀
DeepSWE rates your LLM
Rating exactly how well an AI does on tasks has been an open field. There are benchmarks, but there have been lots of arguments these current early benchmarks are too limited or biased. A new player is on the field and they seem to have discovered that some benchmarks are actually evaluating incorrectly a shocking amount of times.
A startup called Datacurve released a benchmark it says does a much better job. DeepSWE, a 113-task evaluation spanning 91 open-source repositories and five programming languages, produces a dramatically wider spread among the same frontier models.
It’s biggest shock is that it claims many benchmarks aren’t even correct. Their tests cover a pretty impressive and wider range of characteristics such as bigger, more representative tasks. They avoid what they call ‘contamination’ that results from benchmarks that rely on simple Github coding samples that some models simply regugitate vs truely generate. And most damning – they found that some benchmarks verifiers (the part of the code that verifies what the AI built is correct) gives false positive/negative rates from 8-24% of the time.
More benchmarks and more testing is valuable for evaluating models – so hopefully these guys will help push the industry to more scrutiny and reproducible real-world results.
You can even go download it and try it on your own models.
Links:
BMAD and Ralph Wiggum
Do you want to write an app? Don’t know anything? How about something so simple that even Ralph Wiggum could use to generate a working app?
BMAD and the Ralph Wiggum loop (invented by Geoffrey Huntley) are methods that creates an AI loop that first builds something, then tests it. It can help you not only create apps and solutions – but also continually improve them. It launched a whole developer movement.
The existential crisis is real
Vibe coding is causing software engineers to have an existential crisis. What happens when you have an ‘easy’ button that largely spits out things that just work? What are you even doing anymore?