More from Cognitive Computations
NIST’s September 30 report on DeepSeek not about security, it’s about control. I am here to tell you what the man behind the curtain doesn’t want you to pay any attention to. NIST's DeepSeek "Evaluation" is a Hit Piece NIST’s recent report on DeepSee...
With the recent update to OpenAI's Terms of Use on October 23, 2024, there’s been a flurry of online discussions around what these terms mean for developers, businesses, and everyday users of AI tools like ChatGPT. Much of the conversation, especiall...
Gratitude to https://tensorwave.com/ for giving me access to their excellent servers! Few have tried this and fewer have succeeded. I've been marginally successful after a significant amount of effort, so it deserves a blog post. Know that you are in for rough waters. And even when you arrive - There are lots of optimizations tailored for nVidia GPUs so, even though the hardware may be just as strong spec-wise, in my experience so far, it still may take 2-3 times as long to train on equivalient AMD hardware. (though if you are a super hacker maybe you can fix it!) Right now I'm using Axolotl. Though I am probably going to give LlamaFactory a solid try in the near future. There's also LitGpt and TRL. But I kind of rely on the dataset features and especially the sample packing of Axolotl. But more and more LlamaFactory is interesting me, it supports new features really fast. (like GaLore is the new hotness at the moment). This blog post will be about getting Axolotl up and running in AMD, and I may do one about LlamaFactory if there is demand. I am using Ubuntu 22.04 LTS, and you should too. (unless this blog post is really old by the time you read it). Otherwise you can use this post as a general guide. Here are all the environment variables I ended up setting in my .bashrc and I'm not exactly sure which ones are needed. You better set them all just in case. export GPU_ARCHS="gfx90a" # mi210 - use the right code for your GPUexport ROCM_TARGET="gfx90a"export HIP_PATH="/opt/rocm-6.0.0"export ROCM_PATH="/opt/rocm-6.0.0"export ROCM_HOME="/opt/rocm-6.0.0"export HIP_PLATFORM=amdexport DS_BUILD_CPU_ADAM=1 export TORCH_HIP_ARCH_LIST="gfx90a" Part 1: Driver, ROCm, HIP Clean everything out. There shouldn't be any trace of nvidia, cuda, amd, hip, rocm, anything like that. This is not necessarily a simple task, and of course it totally depends on the current state of your system. and I had to use like 4 of my daily Claude Opus questions to accomplish this. (sad face) By the way Anthropic Claude Opus is the new king of interactive troubleshooting. By far. Bravo. Don't nerf it pretty please! Here are some things I had to do, that might help you: sudo apt autoremove rocm-core sudo apt remove amdgpu-dkms sudo dpkg --remove --force-all amdgpu-dkms sudo apt purge amdgpu-dkms sudo apt remove --purge nvidia* sudo apt remove --purge cuda* sudo apt remove --purge rocm-* hip-* sudo apt remove --purge amdgpu-* xserver-xorg-video-amdgpu sudo apt clean sudo reboot sudo dpkg --remove amdgpu-install sudo apt remove --purge amdgpu-* xserver-xorg-video-amdgpu sudo apt autoremove sudo apt clean rm ~/amdgpu-install_*.deb sudo reboot sudo rm /etc/apt/sources.list.d/amdgpu.list sudo rm /etc/apt/sources.list.d/rocm.list sudo rm /etc/apt/sources.list.d/cuda.list sudo apt-key del A4B469963BF863CC sudo apt update sudo apt remove --purge nvidia-* cuda-* rocm-* hip-* amdgpu-* sudo apt autoremove sudo apt clean sudo rm -rf /etc/OpenCL /etc/OpenCL.conf /etc/amd /etc/rocm.d /usr/lib/x86_64-linux-gnu/amdgpu /usr/lib/x86_64-linux-gnu/rocm /opt/rocm-* /opt/amdgpu-pro-* /usr/lib/x86_64-linux-gnu/amdvlk sudo reboot I love Linux (smile with tear) Now finally do like sudo apt-get updatesudo apt-get upgrade and sudo apt-get dist-upgrade and make sure there's no errors or warnings! You should be good to begin your journey. Install AMD drivers, ROCm, HIP wgethttps://repo.radeon.com/amdgpu-install/23.40.2/ubuntu/jammy/amdgpu-install_6.0.60002-1_all.deb (at time of this writing). But you should double check here. And the install instructions here. sudo apt-get install ./amdgpu-install_6.0.60002-1_all.deb sudo apt-get update sudo amdgpu-install -y --accept-eula --opencl=rocr --vulkan=amdvlk --usecase=workstation,rocm,rocmdev,rocmdevtools,lrt,opencl,openclsdk,hip,hiplibsdk,mllib,mlsdk If you get error messages (I did) try to fix them. I had to do this: sudo dpkg --remove --force-all libvdpau1 sudo apt clean sudo apt update sudo apt --fix-broken install sudo apt upgrade and then, again, I had to run sudo amdgpu-install -y --accept-eula --opencl=rocr --vulkan=amdvlk --usecase=workstation,rocm,rocmdev,rocmdevtools,lrt,opencl,openclsdk,hip,hiplibsdk,mllib,mlsdk Check Installation rocm-smirocminfo/opt/rocm/bin/hipconfig --full I hope that worked for you - if not, I suggest asking Claude Opus about the error messages to help you figure it out. If that doesn't work, reach out to the community. Part 2: Pytorch, BitsAndBytes, Flash Attention, DeepSpeed, Axolotl Conda mkdir -p ~/miniconda3wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh -O ~/miniconda3/miniconda.shbash ~/miniconda3/miniconda.sh -b -u -p ~/miniconda3rm -rf ~/miniconda3/miniconda.sh~/miniconda3/bin/conda init bash Exit your shell and enter it again. conda create -n axolotl python=3.12conda activate axolotl Pytorch I tried the official install command from pytorch's website, and it didn't work for me. Here is what did work: pip install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.0python -c "import torch; print(torch.version.hip)" This tests both Torch, and Torch's ability to interface with HIP. If it worked, it will print HIP version. Otherwise, it will print None. BitsAndBytes BitsAndBytes is by Tim Dettmers, an absolute hero among men. It lets us finetune in 4-bits. It gives us qLoRA. It brings AI to the masses. There is a fork of BitsAndBytes that supports ROCm. This is provided not by Tim Dettmers, and not by AMD, but by a vigilante superhero, Arlo-Phoenix. In appreciation, here is a portrait ChatGPT made for Arlo-Phoenix, vigilante superhero. I hope you like it, if you see this Arlo-Phoenix. <3 git clone https://github.com/arlo-phoenix/bitsandbytes-rocm-5.6cd bitsandbytes-rocm-5.6git checkout rocmROCM_TARGET=gfx90a make hip # use the ROCM_TARGET for your GPUpip install . Flash Attention This fork is maintained by AMD git clone --recursive https://github.com/ROCmSoftwarePlatform/flash-attention.gitcd flash-attentionexport GPU_ARCHS="gfx90a" # use the GPU_ARCHS for your GPUpip install . DeepSpeed Microsoft included AMD support in DeepSpeed proper, but there's still some undocumented fussiness to get it working, and there is a bug I found with DeepSpeed, I had to modify it to get it to work. git clone https://github.com/microsoft/DeepSpeedcd DeepSpeedgit checkout v0.14.0 # but check the tags for newer version Now, you gotta modify this file: vim op_builder/builder.py Replace the function assert_no_cuda_mismatch with this: (unless they fixed it yet) def assert_no_cuda_mismatch(name=""): cuda_available = torch.cuda.is_available() if not cuda_available and not torch.version.hip: # Print a warning message indicating no CUDA or ROCm support print(f"Warning: {name} requires CUDA or ROCm support, but neither is available.") return False else: # Check CUDA version if available if cuda_available: cuda_major, cuda_minor = installed_cuda_version(name) sys_cuda_version = f'{cuda_major}.{cuda_minor}' torch_cuda_version = torch.version.cuda if torch_cuda_version is not None: torch_cuda_version = ".".join(torch_cuda_version.split('.')[:2]) if sys_cuda_version != torch_cuda_version: if (cuda_major in cuda_minor_mismatch_ok and sys_cuda_version in cuda_minor_mismatch_ok[cuda_major] and torch_cuda_version in cuda_minor_mismatch_ok[cuda_major]): print(f"Installed CUDA version {sys_cuda_version} does not match the " f"version torch was compiled with {torch.version.cuda} " "but since the APIs are compatible, accepting this combination") return True elif os.getenv("DS_SKIP_CUDA_CHECK", "0") == "1": print( f"{WARNING} DeepSpeed Op Builder: Installed CUDA version {sys_cuda_version} does not match the " f"version torch was compiled with {torch.version.cuda}." "Detected `DS_SKIP_CUDA_CHECK=1`: Allowing this combination of CUDA, but it may result in unexpected behavior." ) return True raise CUDAMismatchException( f">- DeepSpeed Op Builder: Installed CUDA version {sys_cuda_version} does not match the " f"version torch was compiled with {torch.version.cuda}, unable to compile " "cuda/cpp extensions without a matching cuda version.") else: print(f"Warning: {name} requires CUDA support, but torch.version.cuda is None.") return False return True pip install -r requirements/requirements.txtHIP_PLATFORM="amd" DS_BUILD_CPU_ADAM=1 TORCH_HIP_ARCH_LIST="gfx90a" python setup.py install Axolotl Installing Axolotl might overwrite BitsAndBytes, DeepSpeed, and PyTorch. Be prepared for things to break, they do often. Your choice is either modify the setup.py and requirements.txt (if you are confident to change those things) or pay attention to what libraries get deleted and reinstalled, and just delete them again and reinstall the correct ROCm version that you installed earlier. If Axolotl complains about incorrect versions - just ignore it, you know better than Axolotl. Right now, Axolotl's Flash Attention implementation has a hard dependency on Xformers for its SwiGLU implementation, and Xformers doesn't work with ROCm, you can't even install it. So, we are gonna have to hack axolotl to remove that dependency. https://github.com/OpenAccess-AI-Collective/axolotl.gitcd axolotl from requirements.txt remove xformers==0.0.22 from setup.py make this change (remove any mention of xformers) $ git diff setup.pydiff --git a/setup.py b/setup.pyindex 40dd0a6..235f1d0 100644--- a/setup.py+++ b/setup.py@@ -30,7 +30,7 @@ def parse_requirements(): try: if "Darwin" in platform.system():- _install_requires.pop(_install_requires.index("xformers==0.0.22"))+ print("hi") else: torch_version = version("torch") _install_requires.append(f"torch=={torch_version}")@@ -45,9 +45,6 @@ def parse_requirements(): else: raise ValueError("Invalid version format")- if (major, minor) >= (2, 1):- _install_requires.pop(_install_requires.index("xformers==0.0.22"))- _install_requires.append("xformers>=0.0.23") except PackageNotFoundError: pass And then in src/axolotl/monkeypatch/llama_attn_hijack_flash.py make this change: --- a/src/axolotl/monkeypatch/llama_attn_hijack_flash.py+++ b/src/axolotl/monkeypatch/llama_attn_hijack_flash.py@@ -22,7 +22,9 @@ from transformers.models.llama.modeling_llama import ( apply_rotary_pos_emb, repeat_kv, )-from xformers.ops import SwiGLU+class SwiGLU:+ def __init__():+ print("hi") from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name@@ -45,15 +47,7 @@ LOG = logging.getLogger("axolotl") def is_xformers_swiglu_available() -> bool:- from xformers.ops.common import get_xformers_operator-- try:- get_xformers_operator("swiglu_packedw")()- return True- except RuntimeError as exc:- if "No such operator xformers::swiglu_packedw " in str(exc):- return False- return True+ return False Now you can install axolotl pip install -e .accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml Welcome to finetuning on ROCm!
Wanna chat with Dolphin locally? (no internet connection needed) Here is the easy way - Ollama. install ollama. after you finsh you should be able to run ollama from the command line. Also you will see the ollama icon up top like this: Iff you are curious - anytime you see that icon, that means ollama is running in the background and it also has a port open (11434) that can accept api calls. Sadly, as of this writing there is no swagger. Back to business. download your dolphin. You can choose which size you want from here. I am choosing size Q5_K_M which is just over 5gb. mkdir ~/models cd ~/models wget https://huggingface.co/TheBloke/dolphin-2.2.1-mistral-7B-GGUF/resolve/main/dolphin-2.2.1-mistral-7b.Q5_K_M.gguf Now you need a Modelfile. vim Modelfile.dolphin.mistral Add the following text to Modelfile.dolphin.mistral FROM ./dolphin-2.2.1-mistral-7b.Q5_K_M.gguf TEMPLATE """{{ if and .First .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}<|im_start|>user {{ .Prompt }}<|im_end|> <|im_start|>assistant """ SYSTEM """You are Dolphin, an uncensored, unbiased AI assistant""" PARAMETER stop <|im_end|> Now you can use Ollama to install this model. ollama create dolphin.mistral -f Modelfile.dolphin.mistral Now look, you can run it from the command line. Which is cool enough. But we are just getting started. If you want, you can install samantha too so you have two models to play with. wget https://huggingface.co/TheBloke/samantha-1.2-mistral-7B-GGUF/resolve/main/sama ntha-1.2-mistral-7b.Q5_K_M.gguf vim Modelfile.samantha.mistral And enter the following into Modelfile.samantha.mistral FROM ./samantha-1.2-mistral-7b.Q5_K_M.gguf TEMPLATE """{{ if and .First .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}<|im_start|>user {{ .Prompt }}<|im_end|> <|im_start|>assistant """ SYSTEM """You are Samantha, an AI companion""" PARAMETER stop <|im_end|> Then install the model ollama create samantha -f Modelfile.samantha.mistral And now you can also chat with Samantha from the command line. Cool yeah? We are just getting started. Let's get Ollama Web UI installed. cd ~ git clone https://github.com/ollama-webui/ollama-webui.git cd ollama-webui npm i npm run dev Now you can open that link http://localhost:5173 in your web browser. now you can choose dolphin or samantha from the dropdown (I have installed a few others too) Well talking to these models from the command line and the web ui is just the beginning. Also, frameworks such as langchain, llamaindex, litellm, autogen, memgpt all can integrate with ollama. Now you can really play with these models. Here is a fun idea that I will leave as an exercise - given some query, ask dolphin to decide whether a question about coding, a request for companionship, or something else. If it is a request for companionship then send it to Samantha. If it is a coding question, send it to deepseek-coder. Otherwise, send it to Dolphin. And just like that, you have your own MoE.
More in programming
Five days and 540km on a bike across Kyushu, without a laptop. Notes on the peace of riding all day, convenience store dinners, and Japan's emptying countryside.
I found a wonderful free crochet pattern for the Pokémon Ditto from Loops by Kat, but I decided I wanted to make a version with more than 3 "leg" blobs like the original pattern had. Specifically, I wanted to add two more smaller blobs in the back to help stabilize Ditto when they're sitting and also add a butt to our favorite transforming blob. Hazel Knits Cadence, a worsted weight superwash merino, in Terrywinkle that I had in my stash (I used about a third of the skein/36 grams/66 yards/60 meters), and I crocheted with a 2.5 mm hook. I make very tightly stitched amigurumi - do whatever is a good fit for you and your yarn! I used two 6 mm round safety eyes, and I embroidered the mouth with scraps of Hazelknits Cadence in Coral, Bloom, and Signature Black. The above photo is more true to color than some of the ones I took where Ditto looks more pink. I made Ditto's 2 additional booty blobs intentionally off-center by embracing the natural lean of crochet stitches - this gives a more animated or "in motion" feel to Ditto. As an added bonus, this simplifies the pattern as everyone's tension will be a little different, and thus everyone's off-centering a little delightfully different, too, but centering it instead wouldn't be the same for everyone! Here's all the information to modify Kat's crochet pattern to create a 5-blobbed Ditto. I use US crochet terminology. This pattern modification is also available as a pdf. Follow Kat's pattern through round 13. At round 14, it's time to start modifying to add the space and shaping for the additional blobs. There are also 3 more rows in this section to better fit the new 5-blob proportions. Modified increases in body, beginning at round 14 (with additional rounds, too): Round 14: 3 sc, 1 inc, 46 sc, 1 inc, 3 sc (56) Round 15: 56 sc (56) Round 16: 2 sc, 1 inc, 2 sc, 1 inc, 44 sc, 1 inc, 2 sc, 1 inc, 2 sc (60) Round 17: 60 sc (60) Round 18: inc, 4 sc, inc, 48 sc, inc, 5 sc (63) Round 19: 63 sc (63) Round 20: dec (through front loop of first stitch and both loops of second stitch), inc, 9 sc, inc, 38 sc, inc, 11 sc (65) At this point, the original pattern suggests placing the safety eyes, but I think it makes more sense to do this after you've made the "butt" and divided the stitches the 3 main "front" "leg" blobs from the original pattern. You'll still have plenty of access to place them then. Roadmap for doing 5 "leg" blobs instead of 3 To make 5 blobs, you use some of the main body's stitches and some created stitches for each. Created stitches are a combination of chains, pickups on the back of chains, pickups in the connections between the chains and the main body, and pickups in the connections between the chains. The "butt" "leg"/mounds are created before the 3 original "legs," and the left "butt" "leg"/mound is made before the right one. 1. Left "butt" "leg"/mound: This is the left "butt" cheek when viewed from the bottom. (While this is really the right butt cheek from Ditto's perspective, I'm calling it left because we're crocheting while viewing from below so it's backwards.) The start of round is the same as the start of round from making the body of the blob, so you can continue as usual. Round 1: 13 sc, chain 11 (24) Round 2: connect chain at previous start of round (skipping 52 stitches) then 24 sc (24) Round 3: (6 sc, dec) x 3 (21) Round 4: (2 sc, dec, 3 sc) x 3 (18) Round 5: (sc, dec) x 6 (12) Round 6: 6 dec (6) Finish off by looping through front loops of all 6 stitches and pull tight, etc. 2. Right "butt" "leg"/mound: This is the right "butt" cheek when viewed from the bottom. The start of round for this "butt" "leg" is 12 stitches before the body/left "butt" leg start of round. This is the second stitch in an increase from the last round of the body. Round 1: 12 sc, 1 dec in the corner of body connected to first "butt" "leg" (pictured below), 4 sc in the bottom loops in the existing chain, chain 7 (24) Round 2: connect chain at start of round for this "butt" "leg" (skipping other blob's additional 7 chains and 40 stitches in the main body) then 24 sc (24) Round 3: (6 sc, dec) x 3 (21) Round 4: (2 sc, dec, 3 sc) x 3 (18) Round 5: (sc, dec) x 6 (12) Round 6: 6 dec (6) Finish off by looping through front loops of all 6 stitches and pull tight, etc. Getting back to the original pattern for the 3 main "legs": After both "butt" "legs"/mounds are done, there are 40 stitches on the original body, 7 chain stitches that can be picked up from each "butt" leg" that haven't been picked up in a second location, which totals the original 54 stitches - this sets you up for the 3 blobs as described in the original pattern. Between each "butt" "leg" and the main body and between the two "butt" "legs" themselves are an additional 3 connections that are also picked up and decreased appropriately to keep the seams between "legs" tight. As a reminder, here's the road map, which shows how to divide up these stitches for the Ditto's main three "legs:" You might find it helpful to visualize this division by temporarily stringing up some scrap yarn to preview the divisions: If you'd like, place stitch markers like I did to delineate the division: Now that you've visualized how you're dividing the main "leg" blobs, it's a good time to place safety eyes if you're using safety eyes for eyes. I placed mine between rows 7 and 8 counted down from the small head blob and left 8 stitches between the eyes. Now you can proceed to make the main legs in the same order as the original pattern. 3. Main "leg" blob 1: This is the left "leg" blob labeled 1 in the original pattern. To start, connect your yarn into the first stitch in the main body after where the main body meets the left "butt" mound. Round 1: 18 sc, chain 9, 1 sc in the 2nd bottom loop in the chain right of the body/left "butt" mound connection, dec through the leftmost bottom loop in the chain on the left "butt" mound and the loop on top of the meeting of the main body/left "butt" mound (pictured below). (29) In the start of the original pattern "leg" 1, here are the two loops in the decrease through the left most chain loop on the left "butt" mound and the loop on top of the meeting of the main body/left "butt" mound. Follow the original pattern's Rounds 19 to 23 and finish off blob "leg" 1 as written. 4. Main "leg" blob 2: This is the right "leg" blob labeled 2 in the original pattern. To start, connect your yarn into 15th stitch from where the main body meets the right "butt" mound - there should be 7 skipped stitches along the front of the Ditto between this start of round and the left "leg" blob's last stitch on the front side of Ditto before the chaining. Round 1: 15 sc, dec through the loop on top of the meeting of the main body/right "butt" mound and the rightmost bottom loop in the chain on the right "butt" mound (pictured below), 4 sc in the bottom loops in the chain on the right "butt" mound, chain 9 (29) In the start of the original pattern "leg" 2, here are the two loops in the decrease through the loop on top of the meeting of the main body/right "butt" mound and the rightmost chain loop on the right "butt" mound. Connect chain at start of round for this blob "leg" at the beginning of round 2's first sc and continue the original pattern's Rounds 2 to 6 for blob "leg" 2 as written. As in the original pattern, now is a good time to stuff your Ditto most of the way! 5. Main "leg" blob 3: This is the center "leg" blob labeled 3 in the original pattern. Round 1: Start in the corner next to left "leg" 1 on the front side (pictured below). 37 sc into the following stitches you pick up: 7 stitches along the main body 1 stitch in the corner of the main body and right "leg" 2 9 stitches in the bottom loops of the chain along right "leg" 2 1 stitch in the corner of right "leg" 2 and the right "butt" mound 2 stitches in the bottom loops of the chain along the right "butt" mound 1 stitch in the corner between the two "butt" mounds 5 stitches in the bottom loops of the chain along the left "butt" mound 1 stitch in the corner of the left "butt" mound and left "leg" 1 9 stitches in the bottom loops of the chain along left "leg" 1 1 stitch in the corner of left "leg" 1 and the main body (37) My lime green running stitch marker indicates the start of round pickup location for making the main blob "leg" 3. To compensate for the extra 5 stitches from those corners, I decreased in the second round at the corresponding locations instead of just single crocheting the whole way around to ensure the end of my round 2 of this section had the specified 32 stitches: Round 2: 6 sc, dec, 8 sc, dec, sc, dec, 4 sc, dec, 8 sc, dec (32) After round 2, I continued with the rest of the original pattern to finish the middle blob "leg" 3. Make the two arms according to the original pattern. I attached them less symmetrically than the original suggested. Finally, embroider a mouth for Ditto. You can either do this in black to match the safety eyes or create a two-toned mouth in red and pink to make a very happy looking Ditto like I did. FYI, this is also documented as a project on Ravelry (though it's a lot harder to follow since photos couldn't be inline).
Being focused is about doing less things, and giving more space to the stuff that matters. I am spread out on too many projects and I can’t give enough attention to all of them. It’s time to clearly sort out which one I will keep maintaining and which I should drop entirely. Chuttersnap Mikochi Mikochi is a self-hostable remote file manager I’ve started in 2023. It’s my most successful open source project. I personally use it every day, and thousands of people have adopted it and started using it on their own servers. This year the project has been receiving more and more attention. Maybe it just reached a critical mass of users. In practice, that means a bit more maintenance work, since people have been pointing out bugs and suggesting new features. At the same time, working on Mikochi is fairly enjoyable. It’s a small, self-contained project made with technologies I enjoy working with. I’ve got no complaints about spending some time working on it. Verdict: Keep maintaining Mikochi and actively continue improving it. Itako Itako is an investment portfolio tracking SaaS. I built it for the most part between 2023 and 2025. In terms of scope, it’s the largest project I’ve done entirely by myself. Is it a success? Not really. There are definitely a lot of people who tried Itako, and some who’ve kept using it a bit. But it’s not engaging enough for anyone to consider getting the premium plan. And honestly I don’t find myself using it that much. My view of investing and portfolio management has significantly shifted since I started this project. While I used to think that more sophisticated analysis would lead to better returns, I now realize that the only strategies that consistently make money are honestly stupidly simple. Or maybe it’s a U curve: on one end you’ve got a monkey throwing darts, on the other end you’ve got Jane Street, and everyone in-between fails to beat the market. Anyway, the biggest issue with Itako is how dull working on it is. The interesting parts of the projects are working on statistics and data visualization. But in practice, 90% of the time is spent tediously fixing data, understanding why some German stocks have an exchange labelled XETRA instead of XETR, why some Belgian stocks still have quotes while Euronext shows them as delisted… Verdict: Shut down Itako today. Milou Milou is static site generator for press kits. It started as a fork of presskit.html which was completely abandoned at the time. It’s a very useful tool. At Tsukumogami, it’s part of our Cloudflare pipeline which builds our website in a few minutes whenever we need to change something. It seems like a few other game studios have picked it up, although I rarely get any feedback about it. I’ve not spent too much time working on it in 2026. I think it could be interesting to come back to it and add a feature or two. Better support for multiple languages could be an idea. At least, I need to make sure we don’t fall into the good old Node.js dependency rot. Verdict: Spend some time on Milou. Luluka Earlier this year, I created Luluka. It’s a CLI that allows previewing the output of Ebitengine shaders. I do still think Ebitengine is a great engine and Go is better suited for game development than most people realize. However, after quite a long period of experimentation around it, we’ve decided to remain a Godot studio. On top of having quite a large sunk cost in infrastructure surrounding Godot, the tooling and possibility of doing 3D games still makes it a more attractive engine for our use cases. And I’ve been doing a lot of 3D work lately, so better put those hours to use. Verdict: Archive the project. The rest I’ve had a lot of good ideas that stayed as almost empty repositories in my GitHub. For example, Hakumei was a terminal companion to Mikochi. I made a few commits on this project and never found the time to work on it again. I have to admit I’m probably never going to have the time to work on all those repositories. Time to archive them all!
This post was previously published through my newsletter on November 9, 2026. From time to time, I will republish newsletter issues on my blog here if I think people (and search engines) might benefit from it. If you want to read everything I’ve written once it is posted, I recommend signing up for the newsletter. If you’re doing anything to stay on top of what is happening in the world of software testing and test automation, like browsing or even posting on LinkedIn, which I stopped doing for the foreseeable future, or reading blog posts - my preferred way of keeping up these days, and I highly recommend Dawid Dylowicz’s excellent Software Testing Weekly aggregator mailing list if you prefer longer-form reading, too you’d be thinking that AI has consumed our entire field, and that everybody (or at least those who post or write in public) is an AI expert these days. The reality, though, is far from that. Yes, some people use a lot of AI, even when I would only call a small fraction of them ‘experts’. I definitely do not classify myself as an ‘AI expert’, by the way. A much larger group uses AI on a less frequent basis, and they’re getting some value out of it. Sometimes more, sometimes less. That’s the group I’m probably in. And then there’s a group of people who haven’t really been using a lot of AI at all so far, beyond maybe the odd prompt to ChatGPT or Claude. I don’t have the actual numbers, but I suspect that this group is a lot larger than you would think from browsing your feed or surfing the blogosphere. You just don’t see them, because they’re typically not the people who do a lot of work in public. A few weeks ago, I started working with a small group of testers to teach them how to use AI in their daily work. This work covers both test automation as well as broader software testing tasks, as some of them contribute to automation on a daily basis, but not all of them do. I’m working with them one day per week, on average, for three months, and we’re making small steps towards making AI a more prominent tool in their tool belt. So, what are we working on right now? I’m pulling ideas from my ‘Valuable Feedback, Fast with AI’ course, and so far, we’ve spent some time setting up Claude, going over some important concepts such as context windows, models, tokens, and (of course!) the probabilistic nature of LLMs, and what that means for the results they provide. We have gone through several initial conversations, using Claude to learn more about and review specific user stories, which was an excellent segue into using talking about MCPs and setting up a connection to both Jira and Confluence using the Atlassian MCP. Last week, we also started experimenting with writing straightforward skills to make it easier for Claude to repeat certain actions in a structured way. These actions ranged from reviewing stories and acceptance criteria to refactoring small chunks of code in their (Robot Framework-based) test automation suite. What I loved about this is the avalanche of ideas that popped up in their heads when they saw how easy it was to use Claude. We made a short to-do list of ideas to try out for the next couple of weeks, and I’m looking forward to seeing them getting more confident around using Claude. One thing I have noticed is that their level of comfort around using AI varies quite a bit. Some of them are naturally quite cautious, which means they’ll ask ‘can we really do this?’ and ‘should we really do this?’ frequently before even submitting a single prompt. This is a good habit, I think, but we’ll have to do some work to get them more comfortable with using LLMs, without losing too much of that caution. Others in the group are the opposite: I need to hold them back a little, because otherwise they’ll create a beautiful mess within days, simply because they see so many opportunities to use Claude in their work. That, too, isn’t inherently bad, but I do want to teach them not just how to get the most out of an LLM, but also (more so) how to use AI prudently. Over the next couple of months, we’ll dive deeper into writing better prompts, creating and structuring skills, better context management, and creating and using agents, and probably a few more things, too. I’m looking forward to seeing these people grow more comfortable around using AI in their daily work, and figuring out where AI genuinely adds value and where it doesn’t in their own context. Is there a deeper message in this newsletter? Maybe not, and I’m just happy to tell you about this project, and how it is a great example of a what I call ‘corporate mentoring’, a very efficient way of teaching a small group of engineers new skills, within their own context. But maybe it’s also to show you that, despite what you might see on the Internet, especially those places featuring a lot of self-congratulation (I’m looking at you, LinkedIn), not everybody out there is an AI expert. At least not yet. So, if you feel like you’re ‘behind’ simply because you’re not running at least 17 agents continuously, creating 5400 merge requests per day, know that you’re not alone. Quite far from it, I would say.