> RLMs are not agents, nor are they just summarization. The idea of multiple LM calls in a single system is not new — in a broad sense, this is what most agentic scaffolds do. The closest idea we’ve seen in the wild is the ROMA agent that decomposes a problem and runs multiple sub-agents to solve each problem. Another common example is code assistants like Cursor and Claude Code that either summarize or prune context histories as they get longer and longer. These approaches generally view multiple LM calls as decomposition from the perspective of a task or problem. We retain the view that LM calls can be decomposed by the context, and the choice of decomposition should purely be the choice of an LM.
I was intrigued to see this trending, because it seems to contradict what Google has been saying on earnings calls.
It also made me wonder if this reflects conditions for individuals / SMEs rather than large corporate accounts. And I didn’t expect the story to come out of Durban - I would’ve guessed the US.
I haven't been to Durban lately, but my understanding is that the broader SA economy has been under pressure (high unemployment, etc.), and that can hit smaller/local businesses first. So it could simply be a rough patch for your market right now.
That said, if you’re seeing campaigns picking up on other platforms, you might be onto something, at least for your niche.
Looking at your website/content: you’re selling an experience, and this seems like a product that really benefits from strong visual marketing. Make it easy for someone new to "get it" in the first 10 seconds.
Three ideas:
1) Ride local trends: build demos around what’s currently hot in SA/Durban and showcase that with the Magicpods in short, punchy videos.
2) Consider adjacent use cases: beyond magic shows, this could be compelling for advertising, especially at conventions (e.g., ICC). That might be a natural expansion path if event bookings are slowing.
3) Try Airbnb Experiences, or local platforms, like daddysdeals.co.za :-)
I'd loved to see a strong on-device multi-modal Siri + flexibility with shortcuts.
Besides the "best consumer AI story" they could additionally create a strong offering to SMBs with FileMaker + strong foundation models support baked in. Actually rooting for both!
I recently tried to figure out what their offerings currently are. I'm hoping for `efficent but performant AI compute-chips` by Apple ever since they kicked out Nvidia in 2015 (for the ML Models / Exploration parts bellow). It will be interesting to see how good their products will feel in this fast-paced environment and how much legroom (RAM + Compute) will be left non-platform offerings.
To my understanding, they market their ML stack as four layers [1]:
- Platform Intelligence: ready-made OS features (e.g., Writing Tools, Genmoji, Image Playground) that apps can adopt with minimal customization.
- ML-powered APIs: higher-level frameworks for common tasks—on-device Foundation Models (LLM), plus Vision, Natural Language, Translation, Sound Analysis, and Speech; with optional customization via Create ML.
- ML Models (Core ML): ship your own models on-device in Core ML format; convert/optimize from PyTorch/TF via coremltools, and run efficiently across CPU/GPU/Neural Engine (optionally paired with Metal/Accelerate for more control).
- Exploration/Training: Metal-backed PyTorch/JAX for experimentation, plus Apple’s MLX for training/fine-tuning on Apple Silicon using unified memory, with multi-language bindings and models commonly sourced from Hugging Face.
I saw you experiment a lot with python. Do you have a python-focused sandboxed devcontainer setup for Claude Code / Codex you want to share? Or even a full stack setup?
> RLMs are not agents, nor are they just summarization. The idea of multiple LM calls in a single system is not new — in a broad sense, this is what most agentic scaffolds do. The closest idea we’ve seen in the wild is the ROMA agent that decomposes a problem and runs multiple sub-agents to solve each problem. Another common example is code assistants like Cursor and Claude Code that either summarize or prune context histories as they get longer and longer. These approaches generally view multiple LM calls as decomposition from the perspective of a task or problem. We retain the view that LM calls can be decomposed by the context, and the choice of decomposition should purely be the choice of an LM.