jacobr1

9 ms·

91.
by jacobr1 1y ago
These factors are now getting baked into the tools. Primarily via prompt engineering, and what might be called "agentic design," which is how much complexity you put into a single pass vs some hierarchical layering of agents and t
92.
by jacobr1 1y ago
LLMs start to bog down with it at a certain point too. For a couple of my side projects, I decided to let things rip and now worry about code structure at all. After a certain point some of the changes I wanted to make just started either f
93.
by jacobr1 1y ago
No, the definitions are very much up for debate, but there is an actual process here. "Reasoning" in this case means having the model not just produce whatever output is requested directly, but also spend some time writing out its
94.
by jacobr1 1y ago
And you also can just run a windows VM when needed for a few apps if that works for your use case.
95.
by jacobr1 1y ago
Not the full thing - but I use Informed Delivery[1] from the USPS. You get to see a picture of the envelope via email. With a little bit of Multimodal LLM usage I have their email summarized with important mail flagged for me. [1] https:&#
96.
by jacobr1 1y ago
Thanks for sharing regardless. It was a good overview for those less familiar with the material.
97.
by jacobr1 1y ago
I may have missed it ... but were any direct applications to tech writers discussed in this article? Embeddings are fascinating and very important for things like LLMs or semantic search, but the author seems to imply more direct utility.
98.
by jacobr1 1y ago
That is true, in the sense that the 100/200 level classes are covering programming basics in addition to whatever algorithmic theory is being presented. But beyond that that, programs really seem to differ pretty strongly on applied pr
99.
The Policy Puppetry Attack: Novel bypass for major LLMs(hiddenlayer.com)
313 points by jacobr1 1y ago|231 comments
100.
by jacobr1 1y ago
Yep - that is my take. It isn't a weakness of MCP as such, but it a gap for which no good tooling exists afaik.
101.
by jacobr1 1y ago
There is probably a difference here in a few different market segments. You have the AI tooling companies (foundation models, tooling frameworks, vector dbs, etcs...), AI product companies (Agents for Foo), Companies with existing business
102.
by jacobr1 1y ago
The bigger issue is that we need a good why to authenticate more dynamically. In the context of a desktop app (desktop claude, claude code) where you are running an application locally, for your own purposes it can run as your own user. But
103.
by jacobr1 1y ago
Three big places: 1 - Claude Desktop (and some more niche AI chat apps) - you can use MCPs to extend these chat systems today. I use a small number daily. 2 - Code Automation tools - they pretty much all have added MCP. Cursor, Claude Code,
104.
by jacobr1 1y ago
FWIW - that is what claude code does, at least how I use it. It uses the BashTool to call `gh`
105.
by jacobr1 1y ago
Right ... the best way to think about MCP right now flipping the switch from every app/agent developer building their own client wrappers around 3rd party APIs to either the API providers building them or community maintained wrappers.
106.
by jacobr1 1y ago
If computer use gets good enough (it isn't just yet ... but it is getting better _fast_) then it doesn't matter, you'll be able to browse the sites the same way a human does to bypass whatever shit they want to toss in your w
107.
by jacobr1 1y ago
If you give it access to the internet ... it can basically do anything, exfil all your code, receive malicious instructions. The blast radius (presuming it doesn't get out of your sandbox) is limited to loss of whatever your put in (so
108.
by jacobr1 1y ago
In some sense that is actually what MCP is. A way to document APIs and describe how to call them, along with some standardized tooling to expose that documentation and make the calls. MCP hit a sweet spot of just enough abstraction to wrap
109.
by jacobr1 1y ago
Exactly. It is pretty common for APIs to actually signal this too. Headers to show usage limits or rates. Good error codes (429) with actual documentation on backoff timeframes. If you use instrument your service to respect read and respect
110.
by jacobr1 2y ago
It is more subtle than that. It is trained on "everything" so the specific adjectives and words you use in your prompt will cause it to generate very different responses. The fine tuning causes it to prefer by default more reliabl
111.
by jacobr1 2y ago
Are you using a paid version? Do you use web-search? And have you tried alternatives like Claude? I've mostly switched to using Claude these days, with MCPs for websearch and fetching specific remote or local files. It answers question
112.
by jacobr1 2y ago
And the LLM won't be necessary either for the common search case of "dc plane crash" or whatever, but increasingly that will be asked of general assistance agent, and it will dispatch an internet search of some kind, and retu
113.
by jacobr1 2y ago
The broader problem is "cognitive load to understand the code I'm looking at." There are a variety of factors that lead enabling that. This is a very limited example. function isReadyToDoThing(Foo foo) { return foo.re
114.
by jacobr1 2y ago
There is a call-stack depth problem here that is specific to codebases though. For one familiar with the the conventions, key data abstractions (not just data model but convention of how models are structured and relate) and key code abstra
115.
by jacobr1 2y ago
> defining discreet units of AI actions that can be chained into workflows that interact with non-ai services. You got. But that is the interesting part! To make AI useful, beyond basic content generation in a chat context you need inter
116.
by jacobr1 2y ago
One way to think about it is job orchestration. You end up with some kind of DAG of work to execute. If all the work you are doing is based on context from the initiation of the workflow, then theoretically you could do everything in a sing
117.
by jacobr1 2y ago
We've had good success with Black Soldier Flies[1] too. They compost pretty much anything - you can toss in meat/cheese and similar scraps not just veggies. And you get a very high quality protein to feed the chickens. [1] https:
118.
by jacobr1 2y ago
Also consider methods that are using reasoning to potentially dispatch additional searches based on analysis of the returned data
119.
by jacobr1 2y ago
A good example of this was the grand challenge with autonomous vehicles. It spurred all sorts of AI and Robotics related innovation that spun into different directions.
120.
by jacobr1 2y ago
Slightly more complicated because a ton of the abuse comes from IPs located western countries, explicitly to evade fraud and abuse detection. Now you can go after the western owners of those systems (and all the big ones do have have large

More ›