bel8

5 ms·

1.
by bel8 4d ago
I'd love to see these: - DeepSeek V4.1 Flash - Kimi K3 - GLM 5.3 (and flash) - hy4-preview - Grok 4.6 All of these can be acessed using a $10/mo OpenCode Go subscription.
2.
by bel8 4d ago
I'm not. Yes we normalized 1m context window and models tend to hallucinate less. But models have been somewhat stagnant since Opus 4.6/7. And in some regards there were even regressions like Claudeisms that are load bearing.
3.
by bel8 5d ago
Those were just examples. I never said it was the only thing happening. And asking to filter out the bubble you don't happen to like doesn't look like a solution.
4.
by bel8 5d ago
I like HN precisely because it tends to tell me what's happening. GitHub or AWS down? New LLM model? Some Apple event? Some famous tech person passed away? Some language achieved 1.0? I think removing what's currently hot opens a
5.
by bel8 6d ago
their screen are made by Samsung so the tech is known.
6.
by bel8 6d ago
dam I'll have to try that one
7.
by bel8 7d ago
It's an iPhone with a foldable screen made by Samsung. So it's hard to inovate here.
8.
by bel8 7d ago
It's not the worst status symbol. It's useful.
9.
by bel8 7d ago
in 2026? LLM performance, for one.
10.
by bel8 7d ago
Cross-platform is a thing. If anyting, Apple stands to gain from what was already developed for Android. Like https://developer.android.com/phones-tablets-foldables Which targets iOS also: https://kotlinlang.org&
11.
by bel8 8d ago
And it probably takes longer for a human to compile all that info.
12.
by bel8 8d ago
Poisoning LLM? I remember at least one study about it. Perhaps but it would have to be very subtle otherwise they might flag your content and shadowban it from future training. My personal opinion is that trying to make LLMs inefficient is
13.
by bel8 8d ago
I think they had this sitting idle for a while, waiting for their AI division to deliver a capable LLM to power it. Now they have Muse Spark 1.3, which is somewhere between Sonnet and Flable (Opus tier in my experience) in intelligence, but
14.
by bel8 8d ago
yup. GPT Luna tends to keep things objective. Muse Spark 1.3 is also one of the better models in this aspect, for me.
15.
by bel8 8d ago
you're right. ads are already infested with AI videos which means it works for some and those are no longer paying the price of fully manual editing.
16.
by bel8 8d ago
fair but that's also a matter of time, in my opinion. 1) software doesn't commit suicide like some pilots did. 2) software has superhuman reflexes, processing power and are not affected by emotion. 3) software also will eliminate
17.
by bel8 8d ago
> ...but for serious work, it will never be a sure shot job handled entirely by AI. Maybe not entirely handed by AI but when AI starts doing it better than most and in a fraction of the time, for a fraction of the cost, why do you think
18.
by bel8 9d ago
The decision to not contribute when you have infinite resources is a hostile action on itself.
19.
by bel8 10d ago
"Developers developers developers!"
20.
by bel8 10d ago
Regardless, saying it would be too much effort for Apple is laughable at best. Even more so when Apple used to release such docs in the past when it was much less financially capable. They just don't care.
21.
by bel8 14d ago
2legit2quit obviously, given the context.
22.
by bel8 14d ago
The fact that they would rather try to win an internet argument by showing indifference to what others have to say than learn something might hint at traits that keep them unemployed. Harsh but maybe this helps changing their mindset.
23.
by bel8 14d ago
It usually boils down to people trying to convince themselves that keeping their macs hot and with very little ram to spare only to get sub 50 tokens per second on a subpar lobotomized (quantized) model is worth it. And I'm not even co
24.
by bel8 14d ago
> We were speculating that maybe it's because a lot of benches are best-of-n, and should probably be worst-of-n. Thank you. I just changed my opinion on this thanks to you. I agree now, since we tend to execute LLM tasks once instea
25.
by bel8 14d ago
Props to your parent commenter for including context size. Because 131k context window is prohibitively small for my coding workloads so I know a 512GB Mac won't cut it.
26.
by bel8 14d ago
true but if you're actually running k8s and similar workloads, chances are it might eat memory that LLM requires. you'll also notice these articles rarely specify their context window in tokens, because it is small, usually 30k to
27.
by bel8 15d ago
or switch to codex
28.
by bel8 15d ago
it's really not, I see my colleague using https://www.fontlab.com for professional font authoring. I'll ask them about glyphs, got me curious.
29.
by bel8 16d ago
just to clarify, models already ran locally without MLX years before it existed, on non-Apple environments. MLX was just Apple's bridge to what already ran in other hardware.
30.
by bel8 19d ago
I have been using it since it was available in opencode Go plan. It replaced all other models for me. It's much less chatty than DeepSeek V4. And it feels noticeably smarter. I also use Claude at work and GLM 5.3 feels like Opus 4.8 (w

More ›