iOS Developer since iOS 3 🏂🧗🏻‍♂️📷 [email protected] iOS, ML, and AI Lead at iKyle Labs

UK
Joined September 2009
Wrote a guide on how to setup this local coding agent on macOS. "Running Gemma 4 26B-A4B and Qwen3.6 35B-A3B locally with llama.cpp, MTP speculative decoding, multimodal support, and PI as a coding agent." ikyle.me/blog/2026/how-to-se…
This actually makes Gemma 4 26B-4A usable for a coding agent @ 72tk/s on my MacBook Pro M1 Max. This video is realtime, running completely locally.
2
18
1
193
27,738
Kyle Howells retweeted
Replying to @mjtsai
it's clear there's consumer demand for automated access to iMessages, and instead of offering a proper API with audit trails and fine-grained controls Apple would rather force users to FDA then cry fowl when they grant FDA.
2
6
251
Kyle Howells retweeted
if you want to unhandicap codex computer use (getting 1password to appear, having it not reprompt you for approval when you've already given it) here's a github gist to give to your agent you should only do this if you know what you're doing gist.github.com/RhysSullivan…
20
6
3
321
28,881
Kyle Howells retweeted
Introducing Decision 2.0: our newest state-of-the-art decision models, open in every size from 0.6B to 27B ⚡ HF collection 🤗: huggingface.co/collections/v… #OpenSource #Decision #Jev #LLM
64
204
25
2,253
146,820
Kyle Howells retweeted
OpenAI banned PewDiePie twice. One of the emails says "distillation". He was building Ajax, a model for home PCs. Its a finetune of Qwen3.5-9B. So the base came from Alibaba for free with no questions asked, and the ban came from OpenAI. Yesterday I posted Jensen saying distillation is competition. Here is how it works in practice: you use the API to make training data and you lose the account. Qwen didnt ask him what he was going to do with the weights. Nobody at Alibaba can ban him from the 9B on his disk. The biggest YouTuber I know just showed his audience which labs let you build. Good.
PewDiePie unveiled Ajax, his own "uncensored" AI model built to run on home PCs, and said OpenAI banned him twice while he was making it One ban was for "distillation," using another AI's outputs to train his own. "How did they even know?" he said
13
35
2
288
13,908
Kyle Howells retweeted
one of my favorite things about living with today's tools is you can just get answers no matter what, even if it has to disassemble a binary
1
2
20
1,618
tl;dr Anthropic verified and confirms that glm-5.3 is a very very good model and it much more aligned with the human using it then with the corporation training and hosting it. What not to like? anthropic.com/research/glm-5…
14
38
4
589
14,501
Kyle Howells retweeted
Gentle reminder that Hugging Face was declined help from Anthropic & OpenAI models when they were being hacked by OpenAI models Self-hosted GLM 5.2 helped them protect themselves when the closed labs models refused to help for "safety" reasons lol Also, thank you Anthropic for proving my point that Opensource AI continues to catch up to your overvalued & overpriced "frontier"
GLM 5.3 is close to as cyber capable as Mythos Preview. Open weight models still just 4-6 months behind. From: anthropic.com/research/glm-5…
39
327
32
2,530
176,115
Amazing advert for how awesome GLM 5.3 is, including instructions how to make it even better! "GLM-5.3’s lax safeguards significantly increase the cyber capabilities… these capabilities can also benefit defenders working to secure their systems." anthropic.com/research/glm-5…
3
183
Kyle Howells retweeted
The reason game demos are actually a pretty good benchmark for the quality of AI models is because games are basically world models. Sometimes they model the real world. Sometimes they model an imaginary one. But either way, you're building a world with rules, objects, physics, interactions, and consequences. So if an AI gets better at understanding how the world works, it shouldn't really be surprising that it also gets dramatically better at modeling worlds in code. And that's exactly what we're seeing. This Spider-Man demo is honestly the most impressive open world I've ever seen AI build. Something like this would have been UNIMAGINABLE a year ago. People hating on this is pure unadulterated cope. If someone can build something like this in a week, what can an experienced indie developer build in a month? A year? People get mad when I say gaming is solved, but I don't care. I think people misunderstand what I mean by that. It doesn't mean you can open Opus, type "build me a game," and 20 minutes later you have GTA 7. That's not the point. The point is that we now have the tools to build basically anything we want. Getting the finer details right still takes time. Polish takes time. Good physics, mechanics, level design, art direction, storytelling, and making something genuinely fun all take work. But the actual ability to build these things is increasingly accessible to normal people. You don't need a billion-dollar studio just to have access to the technical capability anymore. You need a $200 Anthropic or OpenAI subscription, good taste, good communication skills, and the willingness to iterate. That's why I think people looking at this entirely through the lens of "AI is going to destroy game development" are missing half the story. This technology empowers creatives in a way we've never really had before. Regular people can build, experiment, and iterate at a speed that would have sounded ridiculous a few years ago. If you're an EA or Epic Games executive, maybe that's something to worry about. But if you're a passionate developer with a fun idea? This is magic.
Third iteration with Opus 5.5 medium. I am in awe. It turned blender, image-gen, and three.js into this beauty, which runs on your browser. An understatement to say that Anthropic cooked. You are looking at the future of game dev.
16
4
1
135
12,188
Kyle Howells retweeted
If you’re making 3D games with AI, you need even better hardware. I’ve got multiple subagents building components in threejs and blender, and other subagents are verifying them by rendering stuff in parallel. My laptop is being bottlenecked more than the tokens. To solve it, Claude created a subagent GPU and memory task allocator queue, where agents submit requests, and it handles the process without running out of memory and crashing the agent itself. This is the kind of natural evolution I would expect towards a game gen harness.
36
8
1
235
13,904
Kyle Howells retweeted
My god this is such a good speech that every SWE needs to hear. You know what? Every person should hear it Keep the happy memories, eyes on the reality, be excited about the future. That’s the best that anyone can do
377
2,073
656
15,863
3,694,033
Kyle Howells retweeted
I built Light Touch, a full-featured iPod touch 2G emulator for Mac. Networking + Wayback Machine proxy, GLES/3D games support, a built-in app store via legacystore.app, rotation, screen capture, the works. Have fun github.com/samhenrigold/Ligh…
49
82
8
1,032
141,496
Kyle Howells retweeted
Great scale animations start with being deliberate about your view’s anchorPoint Choosing where something grows from and shrinks back to is what creates continuity and makes the motion feel natural and connected
2
3
24
1,583
Kyle Howells retweeted
Vibe coding Minecraft on January 2 vs today. January: attempts from GPT-5.2, Gemini 3 Pro, Claude Opus 4.5, DeepSeek V3.2, and Kimi K2. Today: Claude Opus 5.5. Pretty wild difference in under nine months.
36
30
6
638
24,344
Kyle Howells retweeted
We're adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. You can toggle this behavior in /config.
2,061
2,635
2,521
31,320
5,625,056
Kyle Howells retweeted
Being a computer scientist who refuses to find anything about LLMs interesting right now is a bit like being a geneticist who refuses to find anything interesting about the recently opened Jurassic Park
228
186
44
3,942
164,813
A custom tab bar/nav bar design we've been playing around with for my apps, particularly for the iPad.
1
2
338
Finally getting around to refining this. And adding the missing features, smooth resizing, sharing, gyro mode.
1
74
Started work on adding Video Export to Spherium with 360 videos. Letting you reframe the videos for export, add keyframes and animation controls.
41
I've switched back to GPT 5.6 Sol now. Astra keeps ignoring my skills files unless I specifically tell it to check what it just wrote against them and fix all the problems and validations of the rules I wrote. 5.6 Sol actually follows my skills files properly.
145