Word and word-like editors are all too complex. Personally I would love some libre quip-like or markdown based renderer with better image and table support
Yeah, I read this and immediately thought “tying arrangement.” But they’re not CHARGING separately for search and AI, so I doubt a tying theory would hold up in court.
Their search engine is now integrated with Ai. So either google would need to block the search engine or the school if they don't want Ai. Take your pick.
A lot of these schools use tablets though. The kids probably use their phones so more likely to be google. Higher grades though use laptops I think so more likely Microsoft then. So far internet explorer is not integrated with Ai but microsoft copilot though is 100% ai and can be used for search.
I'm not talking about Copilot, I'm talking about IE6 back in the 90s, where Microsoft took the "oh woe is us it'll be so hard to undo the deep unnecessary integration we built for this very purpose" tactic.
That happened before some of the readers of this thread were born. A summary for those people, from the fog of distant memories: MS had an Internet Explorer that was not integrated in Windows, it was like any other browser. Then they decided to build it inside Windows. I remember that the Windows desktop got web views like widgets (web view as a name probably did not exist yet) and they called it Active Desktop. I think that it as COM/OLE objects too and the web view was probably one of them. File Explorer itself got an Internet Explorer like skin, probably with single click to navigate directories and open files, as if they were links. Then they claimed that it was impossible to disentangle IE from Windows. That eventually paved the way to the ruling against MS.
When I hear a giant company say that modifying some of its software is impossible I always think about that IE and laugh.
Not really? In the US it kept being bundled by default, they just made it possible to disable the front end; the MSHTML rendering engine was unremovable.
No, because the post-settlement era was actually the era of consolidated IE dominance. Chrome only started because MS, in their (unbroken by the FTC) dominance had sat around doing nothing for the web, which was bad for a web-based company like Google. Chrome also avoided using platform APIs because they wanted to be cross-platform.
Obviously they have at least a back-end search service that isn't tangled up with an LLM. When Gemini says "Searching the web..." it isn't chatbots all the way down. Google just needs to make a human UI for the same search Gemini uses.
Doesn't an enterprise account already use API rates? If that's the case, there's no difference. You'll end up paying less because Pi will typically use less tokens for the same texture of requests.
How is this different than bwrap or srt and others? Im using bwrap to achieve read only everywhere and and write on pwd. Also pi and other coding agents all have sandboxing that work in similar way
read-only everywhere and write only to the current directory is good if you want to prevent accidental damage, such as a coding agent could make if it hallucinated an invalid command, like `rm -rf ~`.
If you want to prevent a damage from a malicious dependency or a prompt injection, you need more robust protection (for example read-only everywhere exposes your ssh keys). You can build this on top of bwrap, but because it is a low level sandbox building block, you would likely end up creating some higher level abstraction on top of bwrap (for example srt and Flatpak are build on top of bwrap).
Drop is an attempt to create such a generic, high level sandboxing tool. I personally prefer to run agents already within a sandbox than to rely on a coding agent runtime to sandbox itself. Especially that by doing so, I can use the same sandboxing tool and config for installing other programs that need isolation, not just for running agents.
If there are no summaries then when context is full messages need to get evicted. If doing so one by one then it would indeed destroy the cache. Of course... maybe the implementation evicted 50% of messages at once, I didnt verify in code
In Jev you pass options in the input and its output just gives some probability for each. Oai structured output just follows a schema. The exact output is still generated and there is no probability
In theory Jev “calibrates” the probabilities, meaning a probability of 20% is optimized to happen near 20% of the time, which as you point out traditional models with schemas do NOT optimize for at all
how do you decide to start a new agent, and when to kill? also... do you use some works-tealing style task board, or otherwise how would the agents get new tasks.
reply