My problem with all of these kinds of things is that memory architecture is the whole point.
If all I needed was common memory space I'd just use symlinks and cloud sync, Obsidian or whatever.
But the context shaping is what makes memory useful in the first place, so just doing one stop shop memory is IMO about as useful as plain old markdown...
With products coming out like this, many claim to do basically the same thing.
How is this better than Claude Code's built in agent orchestrator? Do I need 100 agent types? How do I know the trained agents here are somehow better? Specialization doesn't equate to "better" in every case.
I want to see the light but at this point it feels like these kinds of projects need a better way to benchmark how they are improving on the available state of the art.
It feels like in 2018, a new browser state management tool emerging. Why does this exist?
Aside from it being "instructions for agents", I'm not sure I understand how this isn't just a markdown file that more or less reads like a readme that targets more junior engineers.
I am curious about how this compares to dataview. As a dataview user, I'm not immediately seeing something bases does that dataview doesn't, but I am not a power user.
Garner Health | Data Engineer II | Full-time | NYC Onsite / Hybrid (4 days/week) | https://job-boards.greenhouse.io/garnerhealth/jobs/552040000...
Garner Health (https://getgarner.com) is revolutionizing healthcare economics through advanced doctor performance analytics and innovative incentive models. Our platform is reshaping how organizations access high-quality, affordable care, powering decisions at leading healthcare systems and enterprise clients. We’ve doubled revenue annually for 5 years running, making us the fastest-growing company in our space.
We’re hiring a Data Engineer II to play a pivotal role in building our enterprise-grade data platform from the ground up, ensuring secure data access across our rapidly scaling organization.
Stack: AWS, Snowflake, Argo, dbt, Terraform, Airbyte, JetStreamWork style: Hybrid in NYC. In-office up to 3 days per week.Target compensation: $120,000 - $160,000 + equity
I'll add to the conversation another interesting technique from Chris Voss, which is to use no-oriented questions.
People like to say no. (I'm not sure what this cognitive bias is, but anecdotally I agree.)
So, if you can frame your requests in a way that "no is permission", it will often get a red light a bit easier.
Example: replace "Is this a good idea?" with "is this a bad idea?"
Now, of course "not a bad idea" is not the same thing as "good idea", but it's a lot more likely. Even reading that, I imagine most people would respond more intuitively, because it helps us avoid a commitment we don't necessarily want to adopt.
I think there is some positive effect potential for Apple to let this slide. The broader this network is, the more adoption it receives. P2P as a super-structure has always been a bigger than vendor problem; adoption by any means is likely an allowable tradeoff, especially since Apple doesn't have to do the work here.
Eventually they will capitalize more on the mesh density, rather than crushing the adoption now.
I suspect that, given a reasonable prompt, it would absolutely discard certain phrases or concepts for others. I think it may find it difficult to cross check and synthesize, but "term families" are sort of a core idea of using multi-dimensional embedding. Related terms have low square distances in embeddings. I'm not super well versed on LLMs but I do believe this would be represented in the models.
I've been in the industry for something like 15 years. I've been using LLMs to help me create the stuff I always wanted but never had time to make myself. This is how LLMs can be used by seniors to great effect - not just to cut time off tasks.
Remember - the vast majority of candidates who take the time to do right by your process get zero reward for their effort. You get a reward in the end, so it feels imbalanced. This is true for VERY good candidates, as well.
Not downplaying the amazing progress, but even the video showcases have some weird uncanny valley effects. The winged horse one in particular - the wings and legs morph, the wing on the left disappears and reappears through the tail.
This stuff is a little ways off, but still some amazing effects here. I think it will be a little bit before it is sufficient for production use in any real commercial situation. There's something unsettling about all of the videos generated here.
Use ChatGPT / Claude differently. Instead of prompting it to help you code, prompt it to coach you as a junior engineer, and to stay broad. Focus on principles, etc. Remind you of different pros and cons, and discuss things thoroughly with you. Explore alternatives, etc.
This kind of prompting changes the conversation entirely.
If all I needed was common memory space I'd just use symlinks and cloud sync, Obsidian or whatever.
But the context shaping is what makes memory useful in the first place, so just doing one stop shop memory is IMO about as useful as plain old markdown...