reissbaker

True in the sense of physical security, but I think it's not true here.

Security bugs are just bugs, as Linus likes to say. And the bugs are often also crash or correctness bugs (use-after-free for example usually causes segfaults). For the most part this actually is a productivity boon in the sense of getting higher quality goods out the other side.

To use your apple farm example, it would be like an apple farm where, if the apple rots, it's sometimes just gross but also sometimes it explodes and kills you. Would it be productive to reduce the rot rate of the bomb-apples? I think so!

That's only if you consider security to be unproductive!

If you consider it to be productive work, it's a massive productivity boon happening right now. Of course, product managers might disagree...

No, their report mentions that despite intentionally removing dual-use cybersecurity training data, K3.1 is better than K3 at cybersecurity tasks (both defensive and offensive), because it's just a lot better at coding in general. It's just not as good at cyber as the unrestricted version of K3.1 that they trained initially and didn't release.

This is not rare. Every TCP connection in Go by default enables MPTCP [1]. Simply running a Go program that uses TCP without explicitly disabling this behavior will cause the running kernel to be vulnerable.

1: https://groups.google.com/g/golang-announce/c/PCoQcdxoScw

"One implementation of a feature had a bug, therefore this feature should be turned off on all implementations by default" is probably not going to convince people who think the feature is good.

If you don't want to be able to have clickable hyperlinks, by all means use Konsole — no one is forcing you to use Ghostty. But if you want to know why people like Ghostty's approach vs Konsole's, it's because Ghostty pushes terminal emulator features forward and is low latency. Konsole is low latency and adopts some but not all features that other people invent. It's a noticeable difference and so people who like the feature part of the former like Ghostty/WezTerm/Kitty more than Konsole. If you don't care or dislike that part, and want a more conservative terminal (that is also low latency), Konsole and many others further on the conservative spectrum exist.

Apple Terminal doesn't support:

1. The Kitty keyboard protocol, which makes it impossible to distinguish e.g. Shift-Enter from Enter

2. OSC8 hyperlinks

3. It breaks on APC queries, so you need special handling to not emit them

4. Symbols for Legacy Computing Unicode support

5. The Kitty Terminal Graphics Protocol (or sixel)

6. OSC 9 notifications (or any of the later versions)

and a bunch of other stuff, this is just off the top of my head. For a long time they didn't even support 24-bit color, although they fixed that in Tahoe.

It's totally fine for basic UNIX tooling, but pretty painful if you're trying to make nice interactive TUIs.

Yes, I agree. Ghostty et al don't open-on-click, but on hover they'll show you the URL and there's a keyboard shortcut + click to open. This is much better UX than disabling OSC8.

A couple of annoyances I've had shipping CLI apps to Konsole users:

1. By default, Konsole disables OSC8 link support. Users can enable it manually, but you need to ship fallbacks by default since otherwise the default users will have a broken experience. Ghostty, Kitty, WezTerm, etc all ship working OSC8 support out of the box.

2. No built-in Nerd Font support, so if you want to render icons, you need to tell your users to set up a patched font or else have fallbacks. Ghostty and WezTerm ship built-in Nerd Font fallbacks if your configured font isn't patched.

That being said Konsole is much better than most emulators. iTerm2 is not bad, although it lacks Symbols for Legacy Computing Unicode support (which all of the above have, including Konsole, and which is very useful for drawing prettier terminal UIs). The worst by far is Apple Terminal, truly the IE6 of terminal emulators.

Moonshot has been frontier since this summer. At this point whenever they release a model it's on par with whatever the latest OpenAI/Anthropic model release was.

Not sure I'd put Google there in practice. And while Muse is an impressive jump from behind it's still not quite at the frontier either IMO.

I've also found this to be true of Kimi K3. I still reach for the Astra power tool for especially hard problems, but it needs a lot of cleanup when it's done — whether on hard tasks or easy ones. So unless I need it I prefer the Chinese models, since if they get the job done it doesn't need much post-processing.

(I truly despise the code Fable generates and IMO it has fewer IQ points than Astra so it's not really in my toolbox these days.)

It's a bummer that American open-source is so bad though.

The prediction was a 10x price reduction, so equivocating about "we don't know the size" doesn't save their prediction — a 10x price reduction would imply the model was likely more than 10x smaller in parameter count, and the price factored heavily in their prediction (they predicted its low cost would destroy other competition, since no one else could have such a capable model at that low price). And Haiku/Luna are nowhere near competitive with Astra or Fable, which doubly damns their prediction. I'm surprised you would pooh pooh the equivalence of K3 and original-Fable (which has been widely shown across many different benchmarks and even human preference data), while claiming an equivalence between Haiku and Fable!

Everything starting from mid-2026 is incredibly wrong, which is pretty poor predictive performance considering they published it mid-2025. List of misses:

- In mid-2026, China will be six months behind the released models of the American labs. (In mid-2026, China released Kimi K3, which was equivalent to the best released U.S. model at the time, Claude Fable.)

- In mid-2026, China will nationalize all domestic AI labs and centralize them into a new special economic zone, the "CDZ," created around a nuclear reactor. 80% of all chips in China will be redirected to the CDZ. This uh. Obviously didn't happen.

- In late 2026 OpenAI/Anthropic will release a tiny model that outperforms their previously-best 2026 model (Astra/Fable), that is priced 10x cheaper, and they will market fine-tuning for it. (Lol.)

- AI will increase the stock market by late 2026 by 30% (in fact the S&P 500 is up 13.9% YTD, which is pretty good, but not at all out of the ordinary).

Their predictions for 2027 also seem pretty wrong: e.g. they expect continual learning to be solved in January, and RSI including from a research perspective, not just "optimize CUDA kernels," to also be solved in January, which I suspect no serious lab employee believes anymore. Most lab employees believe RSI will happen (as do I!), of course, but not in two and a half months.

By early 2027 they expect cyberwar with China because China has stolen the weights of a model and smuggled it into the CDZ.

By mid-2027 they expect true superintelligence, where essentially all engineers and researchers are useless, unable to comprehend the advances the AI is able to make and largely unable to even usefully manage or steer AI, who are better at self-managing than a human would be.

Considering their 2025-mid-2026 predictions were basically "progress continues as normal and nothing crazy happens," it's not super surprising that they were more or less right about that. Their 2026/2027 predictions were the spicy ones, and none of them have happened, or are even on track to happening!

At least for RL, you don't need a base model — the rollouts are run in an inference engine with an instruction-tuned model using a chat template. You can start with just that!

Instead of yet another mediocre but fully-made-in-the-West open model (alongside Mistral, Trinity, Poolside, Inkling, etc etc) I'd really love for a Western neloab start the same way Qwen did: by focusing on post-training. Qwen's first release was a Llama 1 finetune [1]! Once they made it useful, they started working their way back in the stack to also do their own pretraining, etc. Starting with pretraining feels like such a waste: there's millions of dollars of crystallized compute and data sitting around in the Chinese model weights. Why not start with one of those, and only work your way back to pretraining once you've released something you can prove is useful?

1: https://en.wikipedia.org/wiki/Qwen

20k tok/sec prefill on B200/B300 isn't particularly noteworthy for medium-sized models like GLM-5.3-Flash, vLLM and SGLang achieve it on a reasonable number of models, especially at NVFP4.

50k tok/sec is pretty impressive though.

But... when you were doing your measurements, were you using the same random book excerpt? If you were potentially getting even partial cache hits for your 50k tok/sec measurement, it would taint the benchmark: pretty much any inference provider running any LLM would be able to hit those numbers.

I feel like you could use an open-source harness like Pi and get 100+% of what these closed APIs offer without getting locked to OpenAI. What do you think is missing from them?

No, OpenAI did not publish how they trained o1, and at the time there was significant misunderstanding and belief in the research community that they were using some kind of Monte-Carlo tree search. DeepSeek figured out GRPO on their own. Similarly, while others invented MoEs, DeepSeek's ultra-sparse variants were extremely novel, to the point where the revelation of how efficient they were to train temporarily collapsed Nvidia's stock.

Regardless I think it's impossible to believe that most LLM research was done by closed labs that don't publish, especially Anthropic (who missed out on and copied two of the largest pieces of important research of the last several years), and that none was done by open labs like DeepSeek, and that the open labs are just copycats. It's quite clear that isn't the case.

It's even worse than that. From the article:

> Open-weights models that don’t have dangerous capabilities are a public good

Oh! And, uh, what's a "dangerous capability" according to Anthropic? Let's see, according to their "Responsible Scaling Policy" [1] document:

- Being able to research energy, robotics, or AI is an unsafe capability

- Additionally, any model that's capable enough to be "used widely" by the government must de facto have unsafe capabilities.

They want to ban pretty much anything open-source that's above cat-level intelligence.

1: https://www.anthropic.com/responsible-scaling-policy

I think it's pretty hard to hold that worldview: Anthropic couldn't ship a reasoning model until they copied DeepSeek R1's homework, and they've all copied DS-style super-sparse MoEs at this point too.

256+: GLM-5.2 (swap with Kimi K3 when it goes open-weight on Monday)

<256: Actually, surprisingly, not a Chinese model but probably Laguna S2.1. The best Chinese model at this size is DeepSeek V4 Flash though

<96: Qwen 3.6 27B

<32: Still Qwen 3.6 27B (NVFP4)

<16: Oof, not sure. Nothing will feel great at this size TBQH without finetuning on a specific task. Pick your poison of tiny Qwen or tiny Gemma (although again Gemma is not Chinese)

The cost to run Kimi is the cost of the GPUs (+ overhead of hiring humans for now to manage it). Kimi K3 does not change the demand curve for LLMs, it only changes the possible suppliers — and they're all competing for the same supply-constrained resource, which is GPUs in datacenters. Regardless of who is serving the model, or how, they're going to need to rent or buy GPUs in datacenters. Hence: this is great for datacenters.

I would assume the opposite is true — with an open-weight Fable-class model, doesn't demand for GPUs go up? Plenty of companies can now look at what Anthropic is offering — high per token costs for a very intelligent model — and do the math, and at some point it makes sense to just rent the GPU yourself and run Kimi on it if you get similar intelligence without paying Anthropic's margins (albeit with high upfront capital cost).

This would drive down Anthropic's margins, but drive up demand for datacenter and GPU capacity. It's not that people would be using fewer GPUs, they'd just shift demand from high priced token vendors to direct GPU rental, which benefits datacenter companies while hurting Anthropic.

Previously they were a distant fourth. They're not going to single-shot catch up to OpenAI or Anthropic, but they moved up the ladder one rung.

In the short term labs are not profitable, although supposedly Anthropic is close. But Amazon was also famously unprofitable for many many years, and then won huge. Current profits or lack thereof are not necessarily important to investors: what's important is they believe in your future potential profits.

In this case, Elon clearly believes much of the economy will be run by AI in the future, and the economic value of a token will rise faster than the cost of generating the token — including the amortized cost of training the model to produce that token. Thus he is building a lab to train models and charge for inference of those models, and — he believes — it will eventually become profitable even if it isn't now.

You may or may not agree with him (and you may or may not agree he's capable of beating Ant/OAI), but current profits aren't a great indicator of whether he believes future profits are attainable. Tesla and SpaceX were also very unprofitable, until they weren't.

Personally I agree with him that there will be massive profits in the future, although I am not as confident in his ability to beat Ant/OAI, at least given his recent difficulties in retaining researchers.

Agreed this sucks. We publish ours here and try to be as transparent as possible: https://synthetic.new/rate-limits

Interesting: I don't see anything in our error logs but we could be missing something (and personally the chat works for me + my unsubscribed test account). If you email us at [email protected] though we should be able to fix anything you're running into!

I'm biased because I run an inference company, https://synthetic.new. That being said I think we're pretty good at serving at GLM-5.2 — and other models, like Kimi K2.7! — and our privacy policy is quite good: zero data retention for prompts and completions on API requests. Our average streaming TPS for GLM-5.2 (aka, tokens after factoring out time-to-first-token, which varies based on geography) is 97tps over the last 24hrs, although it's slightly lower at peak traffic in the mornings PST where it's 50-70 tps. We're also subscription-based which is nicer for coding than e.g. Fireworks which is per-token billing.

Self-promo but you should try our service synthetic.new. We generally have up-to-date open-source LLMs on the sub, and we have GLM-5.2 :) Perf+stability should be wayyy better than zai.

I wonder if this explains why I hear about this more from Europeans than from the SF tech scene. California is at-will employment, so you can fire an employee as easily as a contractor. Ironically this makes companies more willing to hire and retain employees, since they're not worried about getting stuck with a bad one — and most employees aren't bad, and are better for the company than contractors.

The Wikipedia articles say the majority of scholars believe it's based on Aramaic, while a minority of people (primarily non-linguistic-specialists in India) disagree. I think you're the one drawing from bias.

No, modern Hebrew and ancient Hebrew mapped similarly well to the written script — the primary difference between the two is just consonant drift. Both used the same structure of triconsonant roots with affixed patterns, and modern Hebrew morphology is identical to ancient Hebrew (phonemes changed primarily due to consonant drift, but not its structure). Arabic, for example, is similar and similarly well-mapped to its script, as are other Semitic languages that are closely related to ancient Canaanite.