I used to be able to search for parts of names. Or a word that is part of a word in a message. Doesn't work anymore. Only exact matches now, or you don't get anything.

This is google-for-business or whatever they call it now.

Damn, truly we are the products.

They’re pretty good at cross-referencing across multiple formats.

- Serve the model with either Ollama or *LM Studio* (the latter 'hides' their traditional Studio on the download page for the agentic one they are preferring you use).

- For the model, Gemma 4 E4B is a nice combination of performance and speed with a decent context size. Not good enough for complex coding needs, but the Qwen models work okay for that (10GB GPU gets you less than you might think).

- For the interface I'd suggest either OpenCode or Pi.

Personally I run LM Studio and in the UI switch on its built in OpenAI compatible API, then point OpenCode at that and run it alongside (not in) my IDE.

I hope that enshitification calms down a bit whenever the dust has settled on frontier AI companies and we no longer have to force this into every product imaginable. Google used to be pretty good relative to the rest of the big ad tech companies (looking at Meta/Facebook with a scowl...) at giving opt-outs and settings to personalize privacy stuff. This has been a disappointing change to see

And frankly I see the distinction in only number of model parameters vs number of people in cohort it was trained. So 50k-param model trained on 1b emails from 1 million people is ok, encoding/compression loses anything personal. But huge LLMs are closer to lossless compressors that can just spew all my passwords, prescriptions etc verbatim.

The model size also plays a role, since they've become bigger since 2022 (let's use chatgpt as the cutoff date, but it could be around 2020/2021). Modern BERT trained nowadays will keep their lossy compression nature? Or could there be a privacy risk such as it's clear with generative model?

Thus, what are the characteristics of privacy-friendly models? So far, I listed discriminative vs. generative, size, and a lossy transforms (related to size).

My guess is that the pairing of all of these things is due to the fact that Google has likely consolidated everything under the same style of decoder model.

I can still use Google Assistant, I did get an email from Google that they're going to kill it off (the assassination date is 2 months ago, IIRC), but so far Google Assistant still works for me, although it's getting dumber, it failed today when I asked it for "local time in (some city)".

Google's Android Studio Editor has nice enough AI integration, I managed to start vibe-coding a replacement Assistant, that uses on-device speech-to-text, and then I had fun human-coding a parser to understand "set alarm for 8:00" (e.g. if it's 7PM, I want the alarm for 8PM, if it's 10 PM I want it for the next day of course).

I see people try so hard to make this distinction when it really doesn't exist. Transformers are only a little different than long short-term-memory architecture (parallelized, with attention mechanism and no recurrance).

What is the point at which is departed from tradition machine learning? Seems more like people retrofitting a moral decision to not like LLMs and trying to dress it up as a technical distinction.

I found all of the new LLM features to be utterly pointless and incredibly annoying, but the last straw was when my account was abruptly locked recently. For a week I couldn't access mail, maps, photos, digital picture frames for my extended family all went dark, my personal blog ... all shut down with no reason given. No customer support available to contact to resolve google's mistake. A week later, it was suddenly unlocked again. Still no reason given.

Fuck google. You are not their customer; you are their product. Google doesn't care about you any more than Tyson cares about the chickens it sells.

Google has zero customer support even for companies that pay them millions of dollars - their attitude is take it or leave it.

for those who didn't know how to do it:

1. https://mail.google.com/mail/u/0/#settings/general

2. under "Smart features:", toggle off "Turn on smart features in Gmail, Chat, and Meet"

Sorry for the sidetrack, but why is it so common to consider BERT something else than an LLM? It is one of the earliest LLMs so it's nowhere near the size of more recent ones, but even wikipedia classifies it as an LLM:

  Bidirectional encoder representations from transformers (BERT)
  Original author Google AI
  Release October 31, 2018[1]
  Type 

    Large language model
    Transformer
    Foundation model

  License Apache 2.0
  Website arxiv.org/abs/1810.04805 
  Repository github.com/google-research/bert

And in any case it's an autoregressively-trained Transformer, albeit encoder-only. If it's not an LLM, then what is it?

What worries me about the current generation of models is privacy. In another comment, I tried to draw the line where models start to worry me regarding it and discussing it with whimsicalism, we noticed that it's not the generative aspect in particular; model size is probably the main factor. For instance, a larger encoder could make its transform almost lossless and thus more prone to data leakage again, although a decoder-like makes it more susceptible.

Your mileage may vary!