HawtAds

At 60 acres you can probably do a 3000 feet grass strip runway. Get a Cessna and fly yourself, being within flying distance of a tiny airport in this case is pretty awesome. 2.5 hour drive is ~ 1 hour flight time for your C172s.

It's very much real but not necessarily malicious. We track upstream providers pretty closely. Sometimes it's a just matter of a single GPU runtime layer bug/update to break inference outputs. The model weights don't necessarily change/get quantized.

How many requests per second do you have for spam that you are reliably hitting the Luna cache?

It doesn't have to look "cartoony". Harness issue. Your 2026 models are trained on a sufficiently broad set of images that style is almost a solved problem.

> It doesn’t seem like aviation became safer because pilots took risks

That's pretty much the job of test pilots even today. They are usually dual trained pilots and engineers/technicians, and their job is basically to QA new planes. In military aviation especially test pilots are often feeders into the astronaut program (a different but similar type of daredevil).

No, this is just poor management. Japanese 7-11s are ran better because of better culture, products and management. It's not magic.

I think you are overcomplicating things. You have a power envelope, and you have to design your application around it. That's pretty much it. From an engineering perspective, there's no real complexity when it comes to the hardware, software protocols operate one level higher. Your app lives and dies by its power budget.

The benchmark is nice but it's very much biased towards flagships i.e. not very useful in practice if you are trying to ship production mobile apps. Apple historically is extremely stingy when it comes to RAM and they never bothered giving iPads and iPhones more ram until fairly recently (most likely because of ML demands). Your covid era 10th Gen iPads only have 4GB of RAM for the base models.

The Android ecosystem is much more liberal when it comes to RAM because their Dalvik VM JIT (their Java Android Runtime, partially AOT compiled and partially JITed) design is not particularly memory efficient. But the main issue with Android is that their mid/low end (think the Samsung Galaxy A series, the OnePlus Nords, the Motorola Gs etc.) are very inefficient when it comes to single core compute performance compared to iPhones, and it gets worse once you factor in power efficiency. The high end Android flagships running the Snapdragons elites (especially post Oryon acquisition) have no problems matching if not exceeding Apple hardware performance in terms of raw power but they are much more power hungry.

At the end of the day, the current gen of "pocket scale" LLMs are still far from being able to be deployed at scale on mobile. Maybe in another year or two once RAM prices have fallen enough and mobile manufacturers build a lot more matmul and memory circuits into their SoCs instead of a tiny mostly useless "NPU/tensor processor" that doesn't have enough RAM to run anything useful. Current gen production mobile ML still heavily rely on traditional transfer learning like Apple's Core.ML (tiny pre trained models, with perhaps the final layer inference head swapped out).

If you want a good example of production mobile ML, take a look at FUTO keyboard's transformer model for their keystroke prediction/autocorrect which operates in near real-time.

https://github.com/futo-org/android-keyboard/blob/7ebbf5c8b5...

I would say right now 100M-300M params is about the maximum you can do if you want any reasonable level of performance on your "average" non flagship mobile. I would keep to sub 100M for power efficiency sake.

Maybe by traffic but I think Shopify has them beat on valuation (well, it's hard to say since one is a public listed company).

How stable is DuckDB these days? I remember it was very buggy compared to Clickhouse.

I love the article, it's a breath of fresh air compared to the usual bay area centric takes. However there's just one issue:

> The students I want to teach are in the same position, and so are the ones in Nigeria and Bangladesh

It does not take $60 to ship small sub $1 chips from Asia to Nigeria/Bangladesh. These two countries are all on global trade routes (both supply and demand) and the expensive last mile delivery costs are fairly low there too.

Not the first time this sort of thing happened in Canada, fuel starvation is surprisingly common in aviation.

https://en.wikipedia.org/wiki/Gimli_Glider

I would say modern autoland (both GPS based and ILS types) systems are a little better when it comes to handling emergency issues like maximizing range and endurance but the problem with autoland is that they require a lot of things to work perfectly and will refuse to take any risk (autoland systems are not probabilistic like Tesla FSDs, they are mostly automatic traditional control systems). So if the plane will miss the threshold by a few feet the autoland won't even attempt it.

> Is this implying there's no benefit in having the model train on itself?

There are various theories around model collapse when you train on too much AI generated data (that's not for distillation).

Their dinner plate chips are impressive.

Hacker News is very Bay Area/US tech centric where spending a few hundred a month on AI is just pocket change. The weaker AI models with more questionable data retention policies are popular in developing countries. I think the new Facebook muse model will be similarly popular.

You do know em dashes are a normal part of formal prose right? If you go through pre GPT New Yorker and the Atlantic articles, there are plenty of em dashes too.

> Kinda like cultural imperialism but with an ethical spin.

As the old adage goes, US innovates, China imitates, EU regulates.

I would say we are pretty much there.

I wish they would provide an ECharts backend, it's probably one of the most performant and featureful charting libraries you can find.

Interesting, TanStack Charts was also released a couple days ago with a similar design.

https://tanstack.com/charts/latest

> Musk (who freely admits to distilling OpenAI's models) struggle to build a competitive 1T model (Grok 4.5) with massively more compute (Colossus-1 100-250K GPUs, Colossus-2 500K+ GPUs)

Are you sure they are using all of their compute on training? Didn't they rent out a ton to other AI companies?

> LLMs do not care about "elegance" the way human beings do, which is a big advantage.

It's just a matter of time before you can post train it for elegance too. Mathematical proofs in particular can be formally verified automatically which is a big advantage.

Isn't that rather self evident? If you are sampling from a particularly domain constrained vertical, how do you expect the censorship to transfer?

> The distillation data also did not contain any China-sensitive content.

This is a very big disclaimer.

It's like if I generate a dataset focusing exclusively on forestry and arboriculture obviously there won't be any useful censorship, or at least little that can be classified above a statistically significant threshold.

If you want to do a study on something more interesting and useful, do a piece on the various guardrail models of all the major LLM API providers. There are usually both input and output guardrails, and they tend to be almost-black boxes from the model routing point of view.

I don't see what the issue is. Isn't the entire point of LLMs that we can easily translate between languages? What better way to cross check the correctness of a proof than to run it on several different systems simultaneously.

This is just an API level change right? The model itself should be the same I think.

One of the best technical explanations I have heard in a long time.

For model routers, do they have to retrain the routing model every time a new LLM is released?

Not sure about OCR specifically, but the newer (past quarter) vision language models all have a lot of post training on detecting garbled text specifically. You can feed some of the old stable diffusion outputs into a modern model and they can figure out pretty reliably if the text gen is mangled. I think the feature is probably used as part of the RL for the image gen to correct for bad text rendering.

I think OP is referring to the initial sign up based on an invite link, not the database storage itself.

It depends whether you want to check the invite code against the invitee's email. It very much depends on the application. For most use cases, just verifying the code is fine and accept any email addresses that posses it.

Essentially ask yourself this, if you change the UI a bit and allow any user to sign up with any email, and the workspace joining is done via a code they enter after login, is that acceptable for your product design and security posture.