amluto

12 ms·

181.
by amluto 2mo ago
Nice sleuthing, nonsense explanation. An extra TLB flush is never an error. (The CPU is free to flush whenever it feels like doing so.) The error seems to be that somehow a zero-page PTE was present when it shouldn’t have been. This sound
182.
by amluto 2mo ago
I think there’s a difference between what people think they’re aiming at and what they actually aim at. Media DRM makes it harder for the person who can already legally access it to use it or share it in an unapproved manner. Or maybe
183.
by amluto 2mo ago
This doesn’t quite count as “API”, but OpenAI’s roll out of OAuth device code authentication was poor, to say the least.
184.
by amluto 2mo ago
Seems backwards to me. I even bought a 16 Pro because I wanted a nicer phone, and I returned it because I liked it less than the Mini. If there was a new Mini I would have purchased and kept it.
185.
by amluto 2mo ago
The “re-run” part might be rather awkward when your model is poking outside its sandbox to pwn someone else’s system :)
186.
by amluto 2mo ago
I wish Apple had released a Mini Pro. Aim at the high end and give it the top CPU and the better camera. And quit making the thin part thin — just make the whole body as thick as the camera bumps or even 0.5mm thicker to protect the camera
187.
by amluto 2mo ago
One thing I’m curious about: this was apparently a single multi-day run of an agent in an RL harness. What was OpenAI hoping to get out of this run? A single numeric score for RL training? A very long trace to distill into the next model?
188.
by amluto 2mo ago
https://windowsforum.com/security-alerts.84/cve-2025-55241-e... https://orca.security/resources/blog/autowarp-microsoft-azur... https://petri.com/microsoft-azure-automation-cr
189.
by amluto 2mo ago
The OP is pretty clearly not in the US and is not subject to the NEC. And just because the NEC doesn’t say you need to mitigate touch potentials for a specific installation doesn’t mean that you shouldn’t. The NEC is just a bunch of requir
190.
by amluto 2mo ago
I don’t recall the more serious predictions predicting anything worse than what we’ve actually experiencing. There were plenty of predictions about nasty consequences if we experienced X degrees of warming by, say, 2050, and it’s not 2050
191.
by amluto 2mo ago
I stayed in an otherwise lovely flat in northwestern Europe in the summer in an area where the minor heat wave at the time gave daytime temperatures in the mid to high seventies (F) and lows in the sixties. No usable A/C. If there had
192.
by amluto 2mo ago
Windows is useful for running third-party software that, for whatever reason, requires Windows. There is a surprising amount of obnoxiously low quality but nonetheless useful things in this category.
193.
by amluto 2mo ago
The reverse might be worse for the US in the long run. Right now, the US leads China in compute-heavy R&D because the US has access to more compute, and that compute has two components that China has limited ability to produce domestica
194.
by amluto 2mo ago
If you believe Huawei’s marketing, they are doing quite well on their Ascend series with DUV. CXMT is making truly impressive inroads on what is currently a very Large addressable market using DUV.
195.
by amluto 2mo ago
I see no reason to think that CXMT’s yields are poor. Or, for that matter, that Huawei’s yields are a problem.
196.
by amluto 3mo ago
There are many different ways to connect solar panels to the inverter. Usually the exposed metallic parts will be “grounded” (bonded to the ground / protective earth / equipment grounding conductor / whatever your jurisdictio
197.
by amluto 3mo ago
> Speculative decoding is not really worthwhile for sparsely-loaded models. If wonder if you can train a model to optimize this, by trying to make the expert selection sticky across a few tokens, without too much quality loss. Another fu
198.
by amluto 3mo ago
Presumably it’s MoE and only needs to read a small fraction of the weights per token. Bonus points if you can get decent speculative decoding without becoming ALU-limited.
199.
by amluto 3mo ago
> Omitting Azure, which gives some privacy for some $$$ on their models, but not at the level of high-security. If I were ranking third parties on their ability to safely handle my data without compromising it, I would rank Anthropic pre
200.
by amluto 3mo ago
> this is where the more lifting body turning faster with less roll kicks in Can you explain what you mean? If an aircraft follows a given trajectory (center of mass position as a function of time x(t) where x is a 3D position), then it
201.
by amluto 3mo ago
> Specify a maximum roll rate in training and normal-law software and this problem is solved I’m wondering whether there are major airfields with traffic patterns that require roll rates that would exceed whatever maximum is desired for
202.
by amluto 3mo ago
LFP’s flat discharge curve is actually kind of annoying: it makes it quite difficult to measure the state of charge of a cell. BMSes mostly need to track the SoC by counting coulombs, and balancing a series of cells may be challenging unle
203.
by amluto 3mo ago
> Note, too, that a larger lifting body needs to bank less to execute the same turn as a craft with more loading. This is not really helpful for passenger comfort. For a turn to be comfortable, you want the turn to be such that the effec
204.
by amluto 3mo ago
The numbers are easy to run, and the added power consumption is no joke. It’s even worse if your heat pump fails to achieve Carnot efficiency. Never mind that heat pumps can be heavy and may be completely destroyed by even a tiny micromet
205.
by amluto 3mo ago
You need your chip to not completely fail. Latch-up can destroy things. A complete failure of a SERDES may dramatically reduce the utility of the whole system.
206.
by amluto 3mo ago
Also, reserves do not magically refill until their operators decide to pay to refill them, which competes with ordinary non-reserve demand.
207.
by amluto 3mo ago
I use it in a monorepo to break compile-time dependencies. If I need to adjust the size, I spend a minute or two rebuilding.
208.
by amluto 3mo ago
I have a little class called EImpl that is kind of like std::indirect except that it embeds the impl instead of pointing to it. It takes three template parameters: an embedded struct, a size and an alignment. It static_asserts that the emb
209.
by amluto 3mo ago
How would you propose to distill Fable-style input/output pairs without the CoT? If you use them as SFT input, you’ll be trying to train a model to predict the post-reasoning output without any reasoning, and this seem very unlikely to
210.
by amluto 3mo ago
As far as I know, 100% of those Fable interactions would have had encrypted reasoning blocks, so distillation would be distinctly nontrivial even if the data were somehow available.

More ›