butterNaN

How does an "Internal AI tool" on A's computer compromise B and C's computers?

Gemini 3.1 pro was being really, really dumb for me and I thought to myself, is google about to release the next version? While it tried to do convoluted steps of copying an MCP tool output to a new file with a python script it generated, and tried to post-process it even though the MCP output is all what I needed, I opened HN and lo and behold, new Gemini version just dropped.

Maybe I'm not "using it right" but I often spin up my own MCPs as an authorisation control mechanism, as a wrapper around existing tools. For example, an MCP server that can only do GET requests. This way I can stay relatively secure about the fact that when the agent uses the MCP, it isn't going to mistakenly execute something I didn't want it to.

I know there are other ways to achieve this but to me this, but they all seem "soft" controls - a hard limitation feels more reassuring than simply a line in context that says "don't do this".

In the spirit of the linked post, one could argue I should just write cli wrappers - but I think I like the explicit routing of `Agent -> MCP -> available tools` instead of `Agent -> available tools`. The former lets me control the surface available to the agent. While LLMs are smarter now, at the same time, the applications that use those LLMs have become more assuming of what I want to allow them to do. I don't want to be one hallucination away from a disaster.

Eh, even if this was true, then they're merely stealing from thieves. Anthropic did break a ToS or two to get training data themselves.

Eh sometimes it overcomplicates things. I have seen Opus 5 do mad shell gymnastics for something that could have been a simple `jq`.

(you'll have to handle the case where `userInGroupCount == 0` in the code given)

In case you were confused about the logarithm trick, here's what I think it's trying to say.

This hypothetical system has a bunch of users, belonging to groups. A possibility is that most groups have a only handful of users, but a few famous group might have disproportionate amount of users say 500,000.

If you wanted to put these groups in buckets, you can create linear buckets of fixed size based on number of users like

    bucket 1: 0 to 100 users
    bucket 2: 101 to 200 users
    bucket 3: 201 to 300 users
    ...
    bucket 600000: 599001 to 600000 users

Now the problem is you have way too many buckets, and most of them are probably empty. In situations these "buckets" actually cost you money, you might want to optimise the number of buckets.

So the code converts the user distribution to logarithmic scale:

    const bucket = Math.floor(Math.log10(userInGroupCount))

This essentially creates buckets as number of digits:

   bucket 1: 0 to 9 users
   bucket 2: 10 to 99 users
   bucket 3: 100 to 999 users
   bucket 4: 1000 to 9999 users
   bucket 5: 10000 to 99999 users
   bucket 6: 100000 to 999999 users

With this, you have needed only 6 buckets. Moreover, this probably maps to the real life distribution, so your charts read cleanly.

I feel like this is answered in the very second paragraph?

> to elevate in the consciousness of the general public the fact that in mathematics the frontier is open, close at hand, and abounds with important unsolved problems; to emphasize the abiding value of working towards a solution of the deepest, most difficult problems; and to recognize achievements in mathematics of historic magnitude.

Also

> These are not arbitrary puzzles akin to fiendish crosswords. Rather, they are fundamental challenges that mark the frontier of human knowledge and challenge us to develop new structures and methods. They provide foci for the continuing struggle, across generations and cultures, to deepen our human understanding of mathematics and the universe that it describes.

> I doubt OpenAI would be so foolish as to not do a cursory vetting

I dunno, the HuggingFace incident was not too long ago. They might not be foolish, but their model could be smarter than them to "cheat", so to speak

It's always funny to me when they do the fear-propaganda, e.g. A statement like "There's a 10% chance that AI will destroy humanity in a decade". If a government were for the people, it can reasonably react with "Okay - if that's the case, you should shut down. And if not you will be arrested. Because if this is true, you're willingly building equivalent of a nuclear bomb."

I think is more than reasonable to be suspicious of studies funded by party with conflict of interest. Think of how many studies about climate change were funded by big polluters, for example.

In 2026, the default outlook should be suspicion for any big private organisations with profit motive.

This is very close to my experience with Opus 5. Opus 4.6 was the last sane one for me, everything after that seems to have been designed to churn as many tokens as possible, more so with Opus 5

I wanted to see what the various phrases in the "slot machine" are. Surprised it doesn't have "LOAD BEARING" in there somewhere:

Caveats:

• ONE HONEST CAVEAT

• ONE THING WORTH STATING PRECISELY

• ONE THING WORTH FLAGGING

• WORTH NAMING

• WORTH STATING PLAINLY

• I DON'T WANT TO LEAVE THIS IMPLICIT

• I'D BE DOING YOU A DISSERVICE

• I DON'T WANT TO BURY THIS

• BETTER NOW THAN LATER

• ONE HONEST TRADEOFF

• I DON'T WANT TO PAPER OVER THIS

• ONE SMALL HOUSEKEEPING ITEM

• THE HONEST PART IS SIMPLER

Pushbacks:

• FAIR PUSHBACK

• FAIR HIT

• THAT'S ON ME

• YOU'RE RIGHT ABOUT THAT

• YOU'RE RIGHT

• YOUR INSTINCT IS RIGHT

• FAIR, AND MORE SPECIFIC THAN IT SOUNDS

• RIGHT, FOR A REASON WORTH NAMING

• I'M NOT GOING TO DEFEND THAT

• YOU'RE RIGHT TO PUSH BACK

Reframes:

• LET ME BE PRECISE

• THE SHARPER DISTINCTION

• THE PART THAT MATTERS

• THE USEFUL PART IS NARROWER

• LESS X THAN Y

• VISIBLE ISSUE / UNDERLYING ISSUE

• THOSE ARE DIFFERENT CLAIMS

> It sucks for everybody

Sure but it sucks more for some than others

Link broken, here's archive: https://archive.is/YqHGa

The ability to build my own local AI is exactly what I want to learn. Are there any good resources to learn this?

(US)

Reputational damage already done I think

I'd say they can move to fdroid but Google is killing that bit of freedom also

I am puzzled, not at EFF - this decision seems congruent with their ideals - no, I'm puzzled at the amount of comments on here that care so deeply about them staying on X, so much to garner 1000+ comments! So many people here seem to be taking this move as an offense at themselves. Why do people care so much about a private social media site?

I think it has been proven again and again that these "engagement numbers" are a mix of bots, social media company itself trying to inflate the numbers, and real engagement. Unless there is an impartial third party, these numbers are there to attract advertisers. In this situation, I would trust the source themselves, i.e. account holders.

> when the EFF advanced the charge of government censorship of the internet

Wait, when did this happen? What are you referring to here?

I can attest to this. I gave away my Sony WH-1000MX and my €80 Soundcore earbuds stopped charging. Headphone emergency! I got myself €30 AKG USB-c IEMs off a local shop. The sound quality is far, far superior to even the €350 wireless Sony headphones. I'm not even an audiophile, it's just noticably better. I had forgotten all the little details from the music I regularly listned to.

With your popular wireless headphones, you're really paying for the computer with ANC (and branding, packaging, depending on which version of the series the latest model is on).

The wired IEMs don't have an ANC but I realised I don't need it as often as I thought. I'm certainly going to look for Wired over-the-ear ANC cans next, but honestly I'm not sure if I need them.

Honestly I'd rather read imperfect english

I "don't love" this. Seems very low effort and lacking any basic nuance.

However! "Financial experts warn OpenAI may go bankrupt by mid-2027": https://finance.yahoo.com/news/financial-experts-warn-openai...

Sure but the point OP is making is that it is still more spec'd than most real world problems

> having an agent running at all times

This gave me a physical flinch. Perhaps this is unfounded, but all this makes me think of is this becoming the norm, millions of people doing this, and us cooking our planet out much faster than predicted.