niam

It's not just resolve, either. The org must recognize the problem in the first place, which isn't a given. Especially in smaller/govt orgs. Often, recognition of the problem is wrongly tied to how difficult the foremost suggestion is to implement.

And there are orgs for whom any suggestion can itself be so encumbered by uncontrolled red tape and social costs that relatively minor changes become a project.

This title is easy to misinterpret. If I understand correctly: Codex now encrypts sub-agent prompts and hides those prompts from the user.

edit: originally was "Codex starts encrypting prompts, uses cyphertext for inference instead"

I'm sympathetic to Logseq here. My guess is that many things are true but do not sum attractively to an audience right now.

The db probably is the cleaner place for Logseq's note storage, but invites comparison to Obsidian, whose fundamental unit of information is already a document, where Logseq's is more like a bulletpoint. Logseq being open source where Obsidian isn't could maybe blunt the edge of moving to a less-portable format, but a user can also understandably look at the status of two-way sync and see an unsexy "In-progress" and figure that's not good enough. Logseq also carries some baggage wrt instability, and the argument that the db is the pro-stability move unfortunately needs to be proven over time. Overall seems like a rough spot to be in.

I dunno. I like Logseq and wish them a good launch.

Arguing that "at least John is doing [clown thing] in the open" just dilutes whatever leverage John's supporters had against John on that thing.

I find myself unwanting to be on the side of people who willingly give up leverage.

One similarly egregious UX issue on the latest Android is that pressing buttons in the dropdown tray doesn't give any feedback until the action is complete. I can press the "turn on WiFi" button and receive zero haptics or visual indication that the phone registered my tap, for over a second, and THEN it will decide "okay! let's shake the phone and change button color now".

And if you have a tray button that needs to e.g reach over the network to a HomeAssistant instance that needs to itself reach out to some fuckass IoT vendor server, you may as well not expect any sort of feedback before you close the tray.

They're rigorously sourced in that some set of underpaid clerks is tasked with surfacing their best guess as to what material is most relevant to the court, with optional fact checking competing for their time.

https://www.propublica.org/article/supreme-court-errors-are-...

So the sources exist, but whether they're quality is another question.

Zooming out to the original complaint that "A is B" doesn't imply "B is A" in common English, and then further -- to the goal of having an LLM predict tokens that map closely to truth/logic/helpfulness:

I don't think a person speaking plain English in most contexts should be seen as "correct" to answer the question with a non-list answer, even if the question is shaped to expect one, unless there's an established confidence that the shape of the question wasn't made in error.

If someone asked me in real life who "my child" is on stage, and I had multiple children on stage, I would first say that I had multiple children there, rather than choosing one from the set. It would be most helpful for an LLM in my position to do the same, rather than infer that [because Timmy is niam's child, niam's child ought to be Timmy when queried].

Your question arguably doesn't have multiple answers. There's one answer: the set of children (discounting different orders of the set).

The question "Who is her child" has multiple answers because it asks you to deliver a single answer.

"AMD, with nary antecedent aforenotice, has elected to excise"...

practically they work together frequently

Are the originators of these investigations the same? Might that matter?

Crime shouldn't suddenly be ok if you pay taxes on it

Cool. Nobody is making that argument.

Calling an entity that's forbidden from acting on your behalf an "agent" seems funny but maybe it's meant as a catch-all term. Their use of "assistant" seems better for that purpose.

Would you suggest that white blood cells are "not all that good for the body" because they're amped up during a bout of sickness?

Maybe they aren't! But you can't infer that from the mere correlation. Hence that phrase about its relationship to causation.

I'm not sure that your definition of "throwing money away" corresponds to the OP's.

OP uses that phrase to imply the (un)worthiness of spend. You're using it to mean that it doesn't build or maintain equity, which is true almost tautologically but wouldn't be very meaningful unless your audience doesn't understand what it means to rent something.

I like the React model of components being (ideally) a function of state, but I don't touch hooks where possible. And I don't use React itself when I can use the lighter Preact library instead, which provides signals as an escape hatch.

Interesting. Many times I find the opposite case, where my long tail search on Kagi will turn up SOME stuff that's kind of pertinent to the subject, and I'll swap to Google to see if the results are better there, only for it to barely have anything pertinent.

The main issue I've had with Kagi is that using "before:" and "after:" just seems weirder than it does on Google, and will throw in some stuff that's visibly outside the ranges I selected sometimes.

On the presumption that this isn't a joke: em dashes appear in LLM outputs because LLMs were trained on human text which included them organically. It's not as unordinary as memes suggest.

Python does have a huge training set, but I figure lots of that training comes from disciplines where maintainability or system design isn't as heavily incented. Reports, notebooks, dashboards, etc.

My early experiments with LLM Python seemed to give me that impression, but I'm wondering if it's better now or people have other experiences.

I don't understand what makes these "datacenters" if they're distributed across satellites with WAN-esque interconnect.

Are we overloading the term "datacenter"? Or is it not overloaded but somehow able to achieve datacenter-like speeds / (tail) latency even when distributed across satellites?

At some point the metaphor becomes an encumbrance, rather than something helpful

Well put.

Pop science is what it is, but I'm bewildered when the rhetoric around plants "screaming in pain" weasels its way into my conversations. Then it's a game of understanding whether my correspondent sees it as a funny way of straining language, or whether they think it's close to the truth.

Was just reading that headline the other day. Economic darling Japan emerges from the Lost Decades with perfect banking policy.

I incline myself to be more annoyed at the problem than the folks reporting that the problem still exists.

Coincidentally I started toying around with it this week. It's pretty cool. I've known about it at least since Snowden namedropped it, but the main reasons I hadn't tried it before:

- I already isolate workloads between VMs or containers

- Wayland support isn't really there yet without breaking the interop that Qubes provides

- Personal Qubes use cases (e.g banking) overlap with GrapheneOS profiles for me, which I already use. Though Graphene profiles are less ergonomic in that they don't support templating yet.

But I've thrown it on my carry-around Thinkpad to give it a shot and I like it so far.

When have Proton turned their data over to law enforcement without a Swiss court order?

I considered that but I don't see it being very impactful. It presumes a user who cares enough about "their" ChatGPT that they can't move from a particular model provider, but simultaneously does not care enough that model providers themselves have a financial motivation to shoo users onto their newer and more efficient models.

The transition from GPT4 to GPT5 was not well recieved among this crowd -- nevermind that I think this crowd is pretty small (comparatively) to begin with. I just don't imagine you can build a business on that sliver of a sliver, much less one that justifies OpenAI's spending.

It's my understanding that they predate llms, and internet snivel.

I don't discount this as a possibility but my impression is that the OpenAI brand isn't very sticky.

Internet Explorer being pre-installed on Windows devices didn't prevent it from being demolished by newcomer Chrome throughout the 2010s. Now we're looking at a product that's even less integrated, and whose value is exposed through universal interfaces (human language, images, etc.).

If OpenAI succeeds, I imagine that remarkably little of it will have come from the brand. But subtracting the first-mover brand advantage: they can either compete on the frontier, which seems difficult and bears potentially diminishing returns (particularly wrt to distillation); or compete as a commodity, which I imagine cannot justify their valuation/spend.

It seems very uphill of a battle.

The article you're responding to is making specific operational claims about Claude's (basically non-) relevance. I'd be interested to hear if you're directionally correct, but forgive me if I need more details than "but it integrates Claude".

If we're on the subject of tropes: https://theonion.com/report-stating-current-year-still-leadi...

There's so much unexplored space in licenses [...]

Am I wrong that this is orthogonal to "pick a side"? It sounds like you're suggesting that the sides themselves are inappropriately drawn.

I didn't consider that bundling Search & Assistant maybe puts them in a tricky spot among some users who revile LLM features, and others who will utilize them to the cap. To the degree that the former is subsidizing the latter, or costing them customers (probably not a ton): I can see why separating the two offerings makes sense.

Though I'm sympathetic to the users for whom this would basically be a strict downgrade in featureset.