HN user

pftburger

289 karma
Posts0
Comments92
View on HN
No posts found.

The agent is pretending to be a person _for a reason_

The models are trained on people being people. Once you try deviate from that the model performs worse.

A huge tell for this is how well “reasoning” works. Reasoning isn’t some alternate thinking mode, it’s just (sometimes) hidden internal monologues.

It’s easy to anthropomorphise and assume the model is intuiting, but it’s more like it’s hyping it’s self up to do the thing. That said, it’s easy to confuse “being rude to the model” with giving it more tokens to “think”.

I’d be really interested in what a non word based internal monologue could look like. Google played with this a little with the diffusion based codegen stuff. I wonder how trainable a small nonverbal conceptual package could be.

Glaze by Raycast 5 months ago

Honestly Glaze is brilliant.

I assume there is an extensive set of rails for the agent to tie into. (Compare this to asking Claude to green fields an app. Do you use electron? How are notifications handled? Icons? Permissions?)

It springboard off Raycast’s teams feature so well it actually gives it a real reason to exist. You’re empowering the one systems thinker in the group to export their automations to the rest of the group in a way that’s proven to work: small apps that do one thing. (Big apps get complicated, become full time projects that distract from the task at hand)

Fig tried this but it was just for engineers, the value prop was missing, Glaze seems to get this right.

Very nicely done!

Reasonably sure this has less to do with putting data centres in space and more to do with getting them less auditable (see the other news about X getting raided about Grok)

The fundamental issue preventing keyring aperture integration stems from the AirTag’s reliance on inverse-phase magnetic reluctance in the structural substrate. You see, the enclosure maintains a precisely calibrated coefficient offramular expansion. Introducing a penetrative void would destabilize the sinusoidal depleneration required for proper UWB phase conjugation. The resulting spurving bearing misalignment could induce up to 40 millidarkness of signal attenuation. Apple’s engineers attempted to compensate using prefabulated amulite in the magneto-reluctance housing, but this only exacerbated the side-fumbling in the hyperboloid waveform generators. Early prototypes with keyring holes exhibited catastrophic unilateral dingle-arm failure within mere minutes of deployment. Until we develop lotus-o-delta-type bearings capable of withstanding the differential girdle spring modulation, I’m afraid keyring integration remains firmly in the realm of theoretical engineering—right up there with perpetual motion machines and TypeScript projects that compile without any // @ts-ignore comments. The technology simply isn’t there yet.

Can here to say this exactly. Not saying they don’t raise an interesting point but the complete lack of curiosity why a group of experts in simplicity and accessibility decided to take this path is jarring

AI efficiency gains don’t benefit employees, they benefit _employers_, who get more output from the same salary. When you’re salaried, you’re selling 8 hours of time, not units of work. AI that makes you 20% faster doesn’t mean you work 20% fewer hours or get a 20% raise. It means your employer gets 20% more value from the same labor cost.

Marx: workers sell their capacity to work for a fixed period, and any productivity improvements within that time become surplus value captured by capital.

AI tools are just the latest mechanism for extracting more output from the same wage. The real issue isn’t the technology—it’s that employees can’t capture gains from their own efficiency improvements. Until compensation models shift from time-based to outcome-based, every productivity breakthrough just makes us more profitable to employ, not more prosperous ourselves.

It’s the Industrial Revolution all over again and we’re the Luddites

In other news: Man burns down house using 1k old laptop battery (cells)

Thank the powers that be no one will give my neighbours a permit for that.

I wonder if there is a meaningful limit to number of listening zones. I’m imagining a 3d grid of virtual mics in a space, each with an AI behind it

Heck, train the model on the raw sensor data and you get the most awesome conference mics

Vercel’s V0 has been amazing. I lean backend, and the whole mess of styling is a continual struggle. V0 gets it right 90% of the time

But like everything you have to spec it well

Then any top model basically for duplicating work.

“Here is component B, here is component A and test A. Produce a test for B following the same pattern”

Neato. I’m building my own as well but not as a therapist (I pay humans for that, which is what I recommend. Pay for a pro)

My approach is more how can I make a todo list app that’s not pushing hustle but also not ignorable.