HackerTrans
TopNewTrendsCommentsPastAskShowJobs

guiriduro

556 karmajoined 10 anni fa

Submissions

Apriel-H1: Towards Efficient Enterprise Reasoning Models

arxiv.org
1 points·by guiriduro·7 mesi fa·1 comments

comments

guiriduro
·3 giorni fa·discuss
There's potentially a virtuous circle of increased EU tech investment and increased EU tech provider adoption but given the risk aversion sclerosis that affects the european private investment community this does need to be government led. I'd agree the US private debt fuelled $1Tn build out is mostly unsustainable but when you look at what China was able to achieve with govt support on a ˜120bn AI sector investment it behoves the EU to take a stand, the first 100bn is probably the least risky.
guiriduro
·4 giorni fa·discuss
Pretty sure there would be a number of EU states who'd happily pump some money in if the EU were to mandate the ECB to (incentivise intermediary banks to) buy their corresponding long dated sovereign debt cheaply to fund it.
guiriduro
·7 giorni fa·discuss
Do they dupe VC into enormous datacentre and capacity build-out investment? Why yes, actual money going to the AI hypester pick-axe vendors (and early equity dumpers) in that sense, absolutely yes it does.

Is Big AI on track to pay that back with profit from any foreseeable and defensible business model? Different question. I sincerely doubt it.
guiriduro
·6 mesi fa·discuss
For a critical thinking exam, you should be able to show that you can decipher political gobledegook, identify lies and platitiudes, strawman arguments etc(there would be much less of it anyway as in a restricted franchise very few with votes would be swayed by it, which is the point.)

Non-evilness would be shown by answering questions designed to test basic empathy, and also by the lack of recent recorded crimes or misdemenours.
guiriduro
·6 mesi fa·discuss
The diagnosis I largely agree with. To the point I no longer identify as a (small-d) democrat, it is the tyrrany of the mob's choice of (usually reprehensible) representative, who will have lied and pacted with corrupt elites to obtain his/her position, and been nominally elected by the ignorant. Anyone who wants power should by that fact be excluded. The solution however (lottery-election) is absurd. We should instead restrict the franchise by exam to provably non-ignorant, non-evil critical thinkers so that we get representatives who are non-sociopaths that we can respect.
guiriduro
·7 mesi fa·discuss
Apriel-H1-15b-Thinker-SFT uses incremental distillation from Apriel-Nemotron-15B-Thinker, selectively replacing less critical attention layers with linear Mamba blocks to reduce computational complexity while preserving reasoning quality.
guiriduro
·8 mesi fa·discuss
That's exactly the point (ie just prior to distribution) where a simple sanity check should have been run and the config replacement/update pipeline stopped on failure. When they introduced the 200 entry limit memory optimised feature loader it should have been a no-brainer to insert that sanity check in the config production pipeline.
guiriduro
·8 mesi fa·discuss
Come come now. Elon has built Tesla on something closer to $38bn of taxpayer largesse/avoidance according to the Washington Post.0 I doubt many people would fail to create public entities with healthy casino ponzi stock trading if they had umpteen billion $ of govt funny money propping them up. Hell they might even start a rocket business as a sideline...

0. https://www.washingtonpost.com/technology/interactive/2025/e...
guiriduro
·8 mesi fa·discuss
Don't need to. Would it be a big deal to hop on a plane to e.g. Switzerland once a year?
guiriduro
·8 mesi fa·discuss
No. It would be run by a US affiliate using the Microsoft brand, paying royalties to a global company in some other jurisdiction.
guiriduro
·8 mesi fa·discuss
MS could always refocus themselves as a global company (in the legal rather than marketing-only sense), and move their HQ out of the US, then there could be no Trump tantrums affecting other countries, the worse that could happen would be some sanctions on what would then be their in-country US affiliate, with no ability to affect their other global operations whatsoever. Why haven't they followed this approach? Haven't lost enough customers yet?
guiriduro
·8 mesi fa·discuss
Its pretty decent. Decent enough in fact that I can run a Windows 11 ARM install on vmware Fusion on my macbook m4 pro, and it will happily run win arm and x86 binaries (via builtin MS x86 emulation) decently fast and without complaint (we're talking apps, gaming I haven't tried.)
guiriduro
·9 mesi fa·discuss
I decided a Macbook pro M4 pro was the right option for me, 48gb/36 accessible to the gpus with very decent tokens/s throughput for (increasingly impressive) offline midrange open LLM inference and huge battery life. Nothing in the windows world to touch it. But a few windows-only bits of software I occasionally use now run very happily in a Windows 11 arm vm on a free VMware fusion 25H2 on my mac, with a 5 dollar win oem license.
guiriduro
·9 mesi fa·discuss
https://archive.is/54LhN for the ft article.
guiriduro
·10 mesi fa·discuss
The Landrover side is clearly doing work, even if the Jaguar part was doing poorly before this. There are a lot of suppliers also that end up having to furlough if their main client shuts down. And there are second order effects on local economies when that happens, and therefore on HMRC. And you also don't want to lose engineering skills that might be needed for e.g. military crossgrades/upscales. Ultimately its better to have that continue than not, but I also believe the consequences should be significant on the mismanaging equity holders, esp. Tata, and potentially beneficial for taxpayers longer term, which reduces the moral hazard aspect for a better overall solution.
guiriduro
·10 mesi fa·discuss
Forgive my relative inexperience as a newcomer to C#, but doesn't it already provide things like stackalloc, Span and unsafe for these things?
guiriduro
·10 mesi fa·discuss
> It looks like a process model; isolation between programs with a system for inter-process communication, and running within a single process's memory.

Which is better handled by existing mature and simpler abstractions from the actor model, like Akka.net, or maybe Orleans.
guiriduro
·10 mesi fa·discuss
I'm divided, a lot of people internally and from suppliers depend on them for work. Equally they are majority owned by an indian company (Tata) now that also mandated they use their useless in-group IT outsourcer whose secrity appears to be pretty lax. In the end I'd say the govt should make the condition for a loan that they dilute out much of the equity holders / share out the loss - in order for them to step in and prevent a total loss, and so the taxpayer gets some longer term contingent benefit.
guiriduro
·10 mesi fa·discuss
Two things about backdoor layoffs. Mostly its about who. When its corporate dictat, those most likely to leave are those with other options, ie the best talent. So sure a business might save on severances in aggregate, but it doesn't get to decide on who, but simple statistics show it will be the best who move on. So a demoralised and increasingly mediocre workforce is then faced with a much tougher hiring environment with unfillable positions and the downward cycle continues, destroying customer value and reputation to a far greater degree than any temporary layoff savings. All for what exactly, control? Its the C-suites that should be being marched out the door.
guiriduro
·5 anni fa·discuss
Yes, that's why Google doesn't exist and Altavista rules world search.