abrichr

I’ve been working on OpenAdapt: https://github.com/OpenAdaptAI

It’s an open source attempt at making GUI automation reliable enough for things like EHRs, ERPs, Citrix apps, etc.

A lot of agent approaches basically keep looking at the screen and asking a model what to do next. We’ve gone in a different direction. We use models to learn or repair a workflow, but once it works we try to compile it down into something deterministic. On a normal run there can be zero model calls.

Lately I’ve been working on figuring out how to know that an action actually had the intended effect, rather than just that the click/keystroke happened. So workflows have preconditions, postconditions/evidence checks, and halt if the state is ambiguous. We also recently added a way to send those halted decisions to a phone for a human to approve/reject.

ChatGPT Deep Research dug through Taalas' WIPO patent filings and public reporting to piece together a hypothesis. Next Platform notes at least 14 patents filed [1]. The two most relevant:

"Large Parameter Set Computation Accelerator Using Memory with Parameter Encoding" [2]

"Mask Programmable ROM Using Shared Connections" [3]

The "single transistor multiply" could be multiplication by routing, not arithmetic. Patent [2] describes an accelerator where, if weights are 4-bit (16 possible values), you pre-compute all 16 products (input x each possible value) with a shared multiplier bank, then use a hardwired mesh to route the correct result to each weight's location. The abstract says it directly: multiplier circuits produce a set of outputs, readable cells store addresses associated with parameter values, and a selection circuit picks the right output. The per-weight "readable cell" would then just be an access transistor that passes through the right pre-computed product. If that reading is correct, it's consistent with the CEO telling EE Times compute is "fully digital" [4], and explains why 4-bit matters so much: 16 multipliers to broadcast is tractable, 256 (8-bit) is not.

The same patent reportedly describes the connectivity mesh as configurable via top metal masks, referred to as "saving the model in the mask ROM of the system." If so, the base die is identical across models, with only top metal layers changing to encode weights-as-connectivity and dataflow schedule.

Patent [3] covers high-density multibit mask ROM using shared drain and gate connections with mask-programmable vias, possibly how they hit the density for 8B parameters on one 815mm2 die.

If roughly right, some testable predictions: performance very sensitive to quantization bitwidth; near-zero external memory bandwidth dependence; fine-tuning limited to what fits in the SRAM sidecar.

Caveat: the specific implementation details beyond the abstracts are based on Deep Research's analysis of the full patent texts, not my own reading, so could be off. But the abstracts and public descriptions line up well.

[1] https://www.nextplatform.com/2026/02/19/taalas-etches-ai-mod...

[2] https://patents.google.com/patent/WO2025147771A1/en

[3] https://patents.google.com/patent/WO2025217724A1/en

[4] https://www.eetimes.com/taalas-specializes-to-extremes-for-e...

Source? How recently?

You can reach much higher spend through the API (which you can configure `$claude` to use)

> Like, Copilot could watch my daily habits and offer automation for recurring things.

We're working on it at https://github.com/openadaptai/openadapt.

See also: https://audionotch.com/app/tune/

> But that one year of knowledge building, distribution, early customers, then dominates who has control over the cap table for a decade.

Can you recommend any resources for learning how to do this work yourself?

> While the LLMs get to blast through all the fun, easy work at lightning speed, we are then left with all the thankless tasks: testing to ensure existing functionality isn’t broken, clearing out duplicated code, writing documentation, handling deployment and infrastructure, etc.

I’ve found LLMs just as useful for the "thankless" layers (e.g. tests, docs, deployment).

The real failure mode is letting AI flood the repo with half-baked abstractions without a playbook. It's helpful to have the model review the existing code and plan out the approach before writing any new code.

The leverage may be in using LLMs more systematically across the lifecycle, including the grunt work the author says remains human-only.

Corporate income taxes are treated differently than personal income taxes. You absolutely can deduct corporate expenses in Canada.

You might be interested in https://github.com/OpenAdaptAI/OpenAdapt

The "things" you mention may correspond to internal concept representations encoded in the model's weights. See e.g. https://arxiv.org/abs/2206.13289

Interesting, thanks!

Would something like https://github.com/OpenAdaptAI/OmniMCP make sense to include here?

> Twitter was making around $500 billion a quarter before the Apple ad targeting changes

From https://archive.is/rfBcg:

> Advertising revenue was $1.14 billion during the quarter ended Sept. 30 [2021], in line with consensus estimates.

> I just need to verify whichever part I need to use for something else. I can run the code it produces to see if it works. I can look up the reference to see if it exists. I can Google the particular fact to see if it's real. It's really very little effort. And the verification is orders of magnitude easier and faster than coming up with the information in the first place. Which is what makes LLM's so incredibly helpful.

Well put.

Especially this:

> I can run the code it produces to see if it works.

You can get it to generate tests (and easy ways for you to verify correctness).

Thank you both!

Follow-up question: do US employers ever provide assistance with O-1 or E-2? What is considered "a relatively high level of achievement"?

Hi Peter, I have a few questions:

1. I currently do consulting for US clients through a Canadian corporation. If I accept a full-time job in the U.S. (e.g., on a TN or H-1B visa), can I continue consulting for other clients?

   - Does the answer depend on the visa type?  

   - Are there any restrictions on self-employment or side income?

2. Can I set up a corporation (LLC, S-Corp, or C-Corp) in the U.S. while on a work visa?

   - Can I be a shareholder?  

   - Can I take dividends?  

   - Can I actively work in the business?

3. What legal structures allow me to maximize flexibility while working in the U.S.?

   - Would an O-1 visa allow for more flexibility?
  
   - If I move under a TN visa, can I later transition to an E-2 investor visa to run my own business?

4. What are the best options if I want to move to the U.S. while keeping the ability to do high-paid consulting?

   - Should I aim for a Green Card as soon as possible? 
 
   - Is there a specific visa category that would allow this?

Thank you for your help!

> I wondered how soon will I be able to ask the video player "What am I looking at here, describe the equations" and it will OCR the frames, analyze them and explain them to me.

Seems like https://aiscreenshot.app might fit the bill.

Thanks for sharing!

There is an open issue in OpenAdapt to implement COM support: https://github.com/OpenAdaptAI/OpenAdapt/issues/873

This could be a valuable reference.

Have you tried copying the compilation errors back into the prompt? In my experience eventually the result is correct. If not then I shrink the surface area that the model is touching and try again.

Which hackathon?

https://github.com/microsoft/OmniParser/ is trending on Github

> Check out documentation: https://helenaeverleyz.github.io/pocket/

https://helenaeverleyz.github.io/pocket/:

> 404