jkingsman

So if I'm reading this right, it's basically a specialized embedding of a RAG, and then a model in the middle takes the agent's queries and returns responses in a form that is injecting into the main model's input token stream directly i.e. skipping actual text. And the injection format is INCREDIBLY token efficient. Sort of injecting a vibe rather than quotable text. Their results of "good on concepts, less good on quotes/exact text retrieval" seem to intuitively line up with that.

Anyone else read the title and think someone found a way to spot early stock market sentiment shifts by analyzing options patterns?

Oh boy. I wonder how many other companies that have previously been lax about validating that cert revocation is handled correctly by their devices are going to validate the eMudrha cert revocation and discover that they are not, in fact, handling revocation correctly. Yikes.

AO3 rocks, as does the parent organization, the Organization for Transformative Works. They do a lot of pro bono legal work around free speech and copyright (which can be especially sticky around transformative works). AO3 has drawn ire and appreciation alike for their fierce free speech stance (which is, essentially, if it's against the law to exhibit in the state of New York (or against their TOS e.g. not a fanwork), it must come down. Everything -- EVERYTHING -- else stays.)

I also love another of their programs, Fanlore[0], a treasure trove of obscure internet fanfiction works, social groupings, and tropes. Sort of a combination oral history + TVTropes-esque fanfic breakdown.

Fanfiction is a beautiful thing and a great way to stretch creative writing muscles that can be helpful by being a bit less of a blank canvas than needing to invent characters and locations. Of course, a TON of fanfiction does invent, liberally, but I've seen people insist they can't write creatively, then make something beautiful when you ask them to rewrite a scene in their favorite show or book that they were disappointed with and boom, they just wrote their first fix-it fic[1].

Also, praise be for AO3 having no news feed/algorithm, and instead an incredibly intricate tagging system. Makes it easy to find what you want, avoid what you don't (see free speech commitment), and feel liberated from algorithmic exhaustion.

[0]: https://fanlore.org/wiki/Main_Page

[1]: https://fanlore.org/wiki/Fix-it

I struggle with approaching this as I do with many quants, in that there's not really a good measurement of how lobotomized by quantization they are other than "a bit but it's not too bad". Ternary weights immediately make me think "unusably quantized" but I'm presuming that's not the case if someone's standing by this release without massive caveats.

I struggle with deciding whether to take the speed hit on q8 vs q6; I wish it was more standard to benchmark quants or have some way to indicate how much intelligence you're leaving on the table.

I don't mind waiting twice as long for twice as smart; I don't want to be passing up the opportunity of 4x speed-up with 98% intelligence retained. But I have no way to differentiate which is which with quants in any vaguely objective way.

Frustrating, as a hobbyist.

> I really don't get the sense there's any hidden prompt contradicting what's visible

It will happily disclose its entire system prompt (which is interesting in its own right, and worth a read) or pop a reverse shell for you

> I don't see how they could have done a "better" job with this

I generally agree -- the openness is great. However, from experiences both inside and outside of Meta, good execution, a hacker ethic, and transparency ultimately has very little propping it up when money is on the line. In fact, it could be argued that Meta has a shareholder obligation to do profitable things such that even the best intentions can (and usually will) fall in the face of corporate hierarchy and sales numbers.

I think they did a pretty bang up job with Muse (the lack of communication with first-time agent users around how powerfully and confidently they can make horrifying mistakes, and how careful you need to be with prompting, and how even that sometimes isn't enough, notwithstanding).

I also think it will inevitably be used to squeeze profit, and given Meta's history, I think it's almost comical to not assume that will involve violations of the spirit of privacy. (and that's assuming that a proliferation of "it deleted all my files" "it messaged my ex" "it leaked private info" doesn't poison consumer sentiment before it even gets off the ground)

I think it's actually a wink and a nod towards the social media meme of "Le Chaton Fat," a fictional model that is jokingly attributed to Mistral, usually with century-defining benchmarks and unfathomable size.

https://x.com/i/trending/2066326562623127678

Can even have it look for (and enter) sweepstakes within your interests, autonomously, if you give it access to its own email that you don't mind getting blasted with spam.

I'm no evangelist for LLM assistants, but this seems incredibly improbable and represents a failure of MacOS security if so. If full disk access isn't granted, Mac blocks it from the Downloads folder, to say nothing of actually sensitive paths. I would expect a far more likely case of an accidentally granted permission on another device or a permission that was on and then turned off.

Permissionless action is about to skyrocket as an issue, but this particular scenario strikes me as incredibly unlikely. Would be interested to know if Muse can provide more meaningful data provenance/logs.

Scanning iMessage dbs as a passive part of full disk access (and not a messages grant), if true, is a little sketchy, regardless.

Genuinely underrated way to understand circuits. When I was younger and dumber I thought his practice was a bit silly, but man, it really is king.

I will be the first to join you on "most of TSA is security theater"; the baby wipes are glycerine getting flagged by ion mobility spectrometry. The machines are fast because they just look for basic chemical building blocks, not doing full spectrometry to figure out the whole compound. So glycerin, nitrates, etc. set it off (and someone else's sticky glycerin e.g. sunscreen can gunk up the machine for multiple subsequent people).

Ion mobility spectrometry is the tech that powers these, and they're VERY sensitive and VERY prone to false positives because it's just looking for ion signatures ("I smell a nitrate/glycerin/etc."). I'm not sure if they're changing modes on the machines between swabs or something, but the machine is a cue for more in-depth inspection, not "zomg I found a bomb". So, some repeated scanning to rule out a false positive or burn off gunk that's accumulated in the machine i.e. we just got a speck of something on the wand or something is gunking up the scanner as opposed to "someone covered in residue that is scary".

What kind of throughput do you see on what models?

That nearly certainly boots you to secs-per-tok land (as opposed to tok/s). Plausible if you are willing to wait hours to days for responses for simple testing, but not (debatably) "usable".

Contributing to OSM was super cool. I've had an issue where my home snaps to the nearest road, which is in fact fenced off from us, rather than the road that our cul de sac connects to (which was not well marked in OSM). Last week, I went in and in about ten minutes used the visual editor to paint in the driveway, read a quick wiki page on how to correctly label driveways, and then submitted it. I got a pleasant jump scare the next time I doordashed some food and saw the little driveway squiggles I added on the map in the doordash app, and then again for lyft -- such a cool feeling!

Would highly recommend the web editor -- what's proposed here looks super comprehensive but way more difficult to use for a first timer.

I'm not OP, but when I started run-jog-walking one mile, three days a week (which is a bit more than two minute stairs every day, but still asking muscles, joints, and heart to wake up), that my back pain lessened, some niggling knee tightness stopped bothering me, and I noticed an increase in how long I could go doing something moderately strenuous before craving a rest (e.g. moving boxes, or walking with a heavy object, or running after something or someone). It doesn't have to be a massive lifestyle shift -- walking a couple times a week can give similar benefits. But the jump from "basically sedentary" to "loading muscles and joints and heart for ~20 minutes 2-3x a week" is truly ENORMOUS when it comes to our biological systems, and a lot of those systems start running a lot happier with that.

Exercise is not a panacea, and those who say it will fix everything and cure depression and make you perfect are quacks. But, it pains my contrarian ass to say, some light cardio every now and again is much closer to what our bodies are happy with than desk life, such has been my experience.

The math and core experimentation here is beyond my abilities, but what I think I understand is that there are possible deeper patterns of representation that exist in LLMs that are distillations of core conceptual relations in grammar that we can get our heads around in a mathematical sense rather than apparent layer-smeared noise that somehow, un-interpretably (in a meaningful sense), resolve to correct grammar/inferences.

That's pretty cool. I hope I've got that kinda-right.

Not OP, but I use it for personal tasks that are just not worth the claude tokens -- rooting through historical medical records to unify prescription history, super-OCR'ing thousands of PDF pages (i.e. beyond PDF dumping -- vision means it can look at tables, understand tricky things like a continuation of a block quote or aside on the next page, etc.), and when power is cheap I'll just let it noodle on little projects on my data. During my agent's "free time" I give it with a tremendously open ended prompt last week, it did a linguistic analysis of how my texting changes in the lead up to, initiation, maintenance, and ending of romantic relationships.

Query cancels are out of band from the query connection so PGBouncer has to land it on the right server and the right connection. So PGBouncer serves a bogus-but-tracked cancel PID+secret to the client, which then goes to PGBouncer and PGBouncer figures out the right server/process to send the actual PID+secret to. PGBouncer also has to make sure that the connection/PID is still actually running the query; strictly speaking, a client can believe a query is still running when PGBouncer knows it's already complete and the connection has been reused.

Wow, that is a deep level of commitment and learning/exploring; I love it. While I'm sure this is informed by deep preexisting knowledge (to a point -- it's still badass in its own right), I can't help but admire these skills and feel a little inferior about my own.

What a badass level of deep dive.

yeah exactly haha; the threat model implies a level of "you're hosed" that something private I'd say to a friend isn't moving the needle on.

Adding on to this question, do you anticipate the same people capable of tapping phones to think less of you for a dirty joke? The people whose opinion of me would lower for something off-color and the people who possess the ability to wiretap me are a disjoint set lol.

Absolutely. I love building things, but sometimes I want something built. LLM assistance is great for when I want a personal tool, code quality be damned, for a specific purpose, without it taking over a weekend.

I develop an open-source self-hosted client (very mobile friendly) for base-station radios that you want to use from anywhere, and also support MQTT/community observers/bots/webhooks/etc. It spawned out of my need for a daily-driver client that wasn't chained to the radio, and turned into a super fully featured companion client for power users.

The radio API and firmware is open; I have no ire for people who choose to make software closed source so that it can be monetized when there are SO many other options that in many cases supersede the functionality of the closed source option.

https://github.com/jkingsman/Remote-Terminal-for-MeshCore

For sure. This is a straight copy-paste of my prompt which references my architecture, codebase, references folder (God, that's so golden -- .gitignored, but saving the tokens of googling or cloning and just having the codebases we depend on locally is killer) so this is not ready to be copy pasted. However, with the context that this is an interface for a peer-to-peer encrypted radio mesh client (for MeshCore; my code is https://github.com/jkingsman/Remote-Terminal-for-MeshCore), that can maybe give you a mote of context around things that are obviously key (e.g. sending/receiving) or important but not a topline acceptance criteria (message ordering, async radio operations, etc.) to port this and try it out on your codebase.

Prompts: https://gist.github.com/jkingsman/30a61882917c68c114ee28fe5f...

Also, I say "prior to public release" and obviously, this codebase is super publicly released, but that doesn't matter -- what I'm doing in the prompt is priming my agents for a this matters tone. I have no opinions I'd state publicly on the consciousness argument, but I generally dislike deception; in this case, I find declaring this to be our last ditch review before public release puts it in a context state to be more details-oriented than other primers I tried ("this is our final release" led to too many interoperability/future-proof finds; "this codebase has subtle bugs introduced by an adversarial engineer" had too many security false positives or true-but-absurd positives; "please be detailed and carefully dig deep" just wasn't as good. Plus, the "public release" paradigm helped it to do innate classification of "gotta fix before release" vs. "fix soon after" vs. "fix when able" which maps pretty well to my personal taste in the severity of bugs it's found I've evaluated, so I've kept it).

I really identify with this. As an engineer, I really do enjoy building things. However, a lot of times, what I want is a thing that is built. A lot of time, that means I build it, which sometimes I enjoy and sometimes I don't; so many of my half finished projects are things that I still think would be awesome to have but didn't care to invest the time in building.

LLM-driven develop lets me have the thing built without needing to build the thing, and at the same time I get to exercise some ways-to-build I don't use as often (management, spec writing, spec editing, proactive unblocking, etc.). I have no doubt my work with LLMs has strengthened mental muscles that are also be helpful in technical management contexts/senior+principal-level technical work.

Absolutely. I've got a nice multi-paragraph prompt on hunting for subtle bugs, user expectation breaks, crufty/repeated code, useless tests (six tests that actually should be one logical flow; assertions that a ternary is still, indeed, a ternary; etc.), documentation gaps, and a few other bits and bobs.

I sick Opus, GPT5.4, and Gemini on it, have them write their own hitlists, and then have a warden Opus instance go and try to counterprove the findings, and compose a final hitlist for me, then a fresh context instance to go fix the hitlist.

They always find some little niggling thing, or inconsistency, or code organization improvement. They absolutely introduce more churn than is necessary into the codebase, but the things they catch are still a net positive, and I validate each item on the final hitlist (often editing things out if they're being overeager or have found a one in a million bug that's just not worth the fix (lately, one agent keeps getting hung up on "what if the device returns invalid serial output" in which case "yeah, we crash" is a perfectly fine response)).

Absolutely. Effective LLM-driven development means you need to adopt the persona of an intern manager with a big corpus of dev experience. Your job is to enforce effective work-plan design, call out corner cases, proactively resolve ambiguity, demand written specs and call out when they're not followed, understand what is and is not within the agent's ability for a single turn (which is evolving fast!), etc.

It really is incredible. I've been playing through my childhood games on retro handhelds, and recently jumped from <$100 handhelds to a Retroid Pocket Flip, and it's incredible. Been playing WiiU and PS2 games flawlessly at 2x res, and even tackling some lighter Switch games on it.

I think they're saying that OpenSSL is NOT elegant, but that it is successful regardless; hence, code elegance is irrelevant to whether a product is successful or not (and thus that horribly ugly LLM-generated code has a shot at becoming successful).