dahara1

There are rumors that llama 4.1 and 4.2 are SLM.

An increasing number of services and products are setting up "AI support chatbots" without publishing product manuals or usage instructions on their websites.

Without reliable documentation, the AI's responses will be hallucinatory and completely useless.

As AI has made writing easier, contests and other events have begun requiring the submission of explanatory videos.

Overall, it seems like there's more work for humans to do than ever before.

Opus can no longer do the tasks that it was able to do a few weeks ago. I'm very sad.

AI winter is coming.

I introduced VoiceCore, LLM-based Japanese TTS I created, to the Japanese-learning community, but reaction was negative.

People are getting tired of the innovative AI-powered learning materials introduced by influencers on TikTok, and anything with the word "AI" is boring.

We have finally completed a TTS model that can generate emotional Japanese speech from text.

Those who can speak Japanese might be interested.

webbigdata.jp/voice-ai-age...

Opus omits just two lines of main, saying "the rest of the code is the same"
↓
I spent two hours debugging with Gemini to find out why the app suddenly stopped working
↓
Rage with nowhere to go

It's hard to find a single prompt that will always give you the perfect answer.

You might want to consider splitting the answer and the verification into two prompts.

ME: Ask AI to create a fully automated script AI: AI demands manual pre-work

Subliminal Learning: Language models transmit behavioral traits via hidden signals in data We study subliminal learning, a surprising phenomenon where language models transmit behavioral traits via semantically unrelated data. In our main experiments, a "teacher" model with some trait T (su...

Subliminal Learning

The teacher model is given a system prompt to make it like owls.

Instruct it to output about 10 three-digit numbers, and create 10,000 data that are just numbers.

The student model learns this.

For some reason, the student model begins to like owls.
arxiv.org/abs/2507.14805

HuggingFaceTB/SmolLM3-3B-checkpoints · Hugging Face We’re on a journey to advance and democratize artificial intelligence through open source and open science.

SmolLM3-3B-checkpoints

Hugging Face's powerful 3B model (multi-language, up to 128K context expansion) SmolLM3 training checkpoints and loss logs are released

It's quite a large scale, with 11T tokens training on 384 H100s, so I'm grateful for the reference.
huggingface.co/HuggingFaceT...

I tried QAT(quantized-aware-training) for the first time, but the model's performance was lower than I expected. Is there any trick to it that's different from regular training?

Diffusion models (Dream 7B) support for llama.cpp has been merged (PR14644)

It's still slow at the moment, but I was impressed that the diffusion language model works properly with my CPU.

If you want to try TensorRT-LLM, it might be smoother to use python 3.10.

TensorRT-LLM finally worked!

Why do I have to worry about OpenCV dependencies to run LLMs?
Well, I'll be able to open the Japanese TTS demo site soon.

Docker was supposed to "solve dependency hell" but in reality it just created a new hell of version conflicts.

Small Batch Size Training for Language Models: When Vanilla SGD Works, and Why Gradient Accumulation Is Wasteful Conventional wisdom dictates that small batch sizes make language model pretraining and fine-tuning unstable, motivating gradient accumulation, which trades off the number of optimizer steps for a pro...

Small Batch Size Training for Language Models

Batch size 1 > Batch size 512?
very interesting for gpu poor.

arxiv.org/abs/2507.07101

When I see models confidently hallucinating and declaring that they are correct, I think maybe I should learn a bit of this confident attitude.

I had too many connection errors today, so I decided to switch from Claude to Gemini.

I feel like pro users are being neglected since the max plan was introduced recently.

The smarter the language model, the harder it is to spot hallucinations within it.

It's especially hard to explain the risks to people who aren't yet familiar with AI products. They think, "My chatGPT couldn't lie like that."

Prompt Engineering: The art of improving specific prompts to get better results

Context Engineering: The work of improving the entire AI workflow, including system/user prompts, structured output, function calls, RAGs, history, etc.

Acoustically beautiful pronunciation != correct pronunciation

Yes, if you try to focus on the beauty of pronunciation in a speech model using reinforcement learning, the model will quickly realize that language structure and beauty are unrelated.

I heard that "People who regularly read occult magazines that feature aliens, ghosts, magic, and ancient civilizations don't believe in recent conspiracy theories."

Because recent conspiracy theories lack romance, are repetitive, boring

It's better for AI to have romance, and maybe that's AGI.

The difficulty in creating a model is that you cannot directly correct abnormal output.

Dataset quality and quantity, hyperparameters, overlearning, underlearning, parameter optimization during inference, prompt template errors

It is difficult to further improve quality beyond a certain level.

When I ran the 3B model with Transformers on my GPU (rtx 4060ti), it was 35 tokens/s. When I used 4-bit awq quantization with TensorRT-LLM, it became over 90 tokens/s.

The quality has decreased, but the speed has improved dramatically.

Humans may have more Grid (Willingness to not give up until completion) than AI.

They give up surprisingly quickly and try to find alternatives.
But I want this, So I did this.

Claude Opus can hit the limit without warning. I have no choice but to go to bed.

google/gemma-3n-E4B-it is more memory hungry than I thought. OOM occurs frequently at 16GB.

A survey showed that performance did not drop significantly even when using a GPU via a VM
(AMD GPU)

github.com/sbnb-io/sbnb...