estebarb

I really do not get the obsession with translation, particularly bad translations and translating by default. Most people who browse websites in languages other than the official local one are fluent in that other language.

Google Chrome keeps insisting that Spanish is Galician. PowerPoint insists on changing the spell checker back to the wrong language and ignores setting the language for the whole presentation. On Linux, ChatGPT insists it must enable the spell checker in the local system language, so everything is always red.

Privacy aside, a lot of software seems to be built assuming people cannot be multilingual.

I dislike this approach. In my opinion it is much useful a 1x/2x billing based on hour like Deepseek does. Being unable to use it at the hours I need it makes me want to remove the service, not upgrading it.

Now, if I could afford 4...

That is how you get a hardware provider to stop selling you stuff

Personally I have been using Deepseek V4 flash and it is enough for my needs. You can use OpenCode Zen/Go and also get access to many other models, including OpenAI ones. Writing purpose specific agents can make cheaper models close the intelligence gap with the most expensive ones

I'm suspecting it is Deepseek flash-vision-exp V4

What happens if we train models (GPT or human students) using the outputs of a model with text havingbthose watermarks? Is there something preventing the watermark from being learnable?

I prefer V4 flash, but Luna is ok and included in the OpenAI plan I'm already paying, so... that is the reason I use it.

I gave DeepSeek $50 around june, and I haven't been able to exhaust them yet. The model is super cheap and more than enough for my needs.

I'm my opinion, Flash V4 is less pedantic than OpenAI models, less prone to unsolicited prescriptions and less prone to "helpfully" reinterpreting my instructions (wrongly, of course).

I practically switched to doing everything with Luna or DeepSeek V4 flash. I haven't feel the need for the more expensive models.

Language distribution shifts. Eventually people will start adopting the distribution used by LLMs, making classification harder.

Also, this doesn't even consider the case where people use LLMs to translate their original works. Or people that use it for spelling/grammar checks.

Personally, I believe these checkers do more harm than good. Any false positive can ruin someones life.

This will cause cognitive debt anyway. As mentioned in https://arxiv.org/pdf/2509.21972v1: "When students rely on these outputs as a substitute for their own reasoning or critical engagement, the learning process is fundamentally compromised. Genuine learning requires the active construction of meaning, integration of knowledge, and reflective engagement with content. These processes cannot occur through passive consumption of syntactically correct but semantically hollow responses. Without this deeper cognitive work, learners risk mistaking linguistic fluency for understanding, thereby undermining the very goals of education".

Personally, I don't think we will ever be able to reconcile using LLMs and cognitive debt. Even before LLMs we were aware if it: we knew people moving to managerial/PM roles eventually get their coding skills rusted. Well, now we are all in those managerial roles...

Oh wow, even scammers care about usability and their employees' well-being. What's the excuse for bad UX in internal company software?

I'm really not sure how this would work. I don't know how the ACM works, but in IEEE you would have to give them your publishing rights. However, training a LLM is not publishing by itself, it is a derivative work? Any way, at this point authors should be entitled to monetary compensation, not the publisher. The deal is totally different.

Session cookies do not require a banner.

Moving money around is harder than it seems

I don't understand why almost always the AI doomerism needs to give embodiment the AI. Real life is not a movie needing visual effects.

AI doesn't need a body of itself to became dangerous. A politician could get into ChatGPT induced psychosis, that doesn't require a physical body, and it could happen now or even 3 years ago... I'm sure it has already happen...

Other pointed risks are interesting because they are basically the risk of empires loosing their tributary states. The world has always seen war mongering states trying to preserve their abusive advantages, humans doesn't need AI for that.

Overall, I would be far more concerned about some really dangerous humans than AI itself.

I tried the smallest ome, the only available, and got an error:

``` Could not load.

Error: failed to call OrtRun(). ERROR_CODE: 1, ERROR_MESSAGE: Non-zero status code returned while running GroupQueryAttention node. Name:'/model/layers.0/attn/GroupQueryAttention' Status Message: Failed to create a WebGPU compute pipeline: A valid external Instance reference no longer exists. ```

Real programmers use butterflies (mandatory xkcd reference) https://xkcd.com/378/

But what users prefer? Given this is for marketing, which results produce more conversions? From the examples shown, personally I strongly preferred Claude Opus in all cases.

"Uniqlo x Akamai sells another design of shirt in the same range which is plainly incomplete"

Imagine having to return a t-shirt because that malfunction!

— I don't understand why are you returning this, was the size wrong or you didn't like it?

— No, there is a syntax error at line 37 that makes it impossible to run, and I'm concerned people on the street may think I promote unsafe bash scripting.

Hash functions necesarily have collisions. Also, it is perfectly possible to introduce bugs in the hash function (hash inputs, hash function itself) that allows cross account contamination.

That sounds very similar to what we know in self-supervised learning as representation collapse. I wonder if we could copy some of the anti-collapse mechanisms from SSL into GPT... after all, they are ways to increment the differential entropy. However, I'm not sure if it could be useful after all: any pure function cannot produce more entropy than the entropy it receives... and natural language as text has much less entropy than other domains... [edit: typos]

This is interesting. Adam Smith said "People of the same trade seldom meet together, even for merriment and diversion, but the conversation ends in a conspiracy against the public, or in some contrivance to raise prices."

The annoying part is that, as the same Adam Smith says, regulating industries would end up enforcing such assemblies, reinforcing the problem... after all, industries can share information via the market itself...

And proposed solutions end up being controversial: employees ownership, open source, paying taxes over stocks ownership... or just hoping that colluders will be broken by a randomly ocurring incumbent...

I really miss the crimson red. New one makes me think they are mourning someone.

Often, but not always. For my thesis, I ended up with a section related to porn. ChatGPT simply refused to spell-check that section. Also, recently I wrote a comment on HN about different subsets of English being easier to learn for native Spanish or German speakers, and the Samsung AI spellchecker refused to review it because it was considered "inappropriate content."

I had that discussion with my high school English teacher. We used USA oriented books and they often introduce "advanced vocabulary" which should have been trivial for Spanish speakers, or any latin language speaker for the matter.

I suppose they evaluate difficulty based on origin of the word. If you already know German or Spanish you may have a head start when learning English, but on a different subset of it.

I can do high level thinking for around 6 hours with just two scrambled eggs and a cup of coffee.

What I need is something to prevent me from context drift. /starts googling how many scrambled eggs are equivalent to the energy consumed by a data center. Google how many chickens are in the world.../

Simply Ruby on Rails. I maintain 3 markdown documents with system design, implementation plan and use cases in repo. Then tell it those files exist and go implement X feature (from implementation plan). These documents, plus AGENTS.md, declare completion criteria, which includes full code coverage both with system and controller tests.

Usually I don't tell it to implement something adhoc, I first implement it in the documents first. LLMs are quite good to keep those documents in sync.

A good part of the implementation plan is that it keeps the LLM on track. With it, the LLM can understand why something must not be done yet, so it includes less unsolicited functionality. My workflow surely can be improved, but it has worked well for me.

In not sure about the actual costs, because I started using the same subscription for document parsing. But even then, I used less than $10 in may.

The F35 program is essential! When the USA will finally conquer free healthcare?

I'm not sure that is really the case, or relevant in practice. I have been using OpenCode with DeepSeek lately (regular coding). For instance, today I got 120 million input tokens hitting cache, vs just 2.59million missing cache.