ChatGPT, Claude, Grok: Your chats end up with trackers

[Redaktion]

ChatGPT, Claude, Gemini, Grok and Perplexity pass chat links, titles and sometimes entire prompts to trackers from Google, Meta or TikTok, according to a study of nine AI assistants by researchers in Madrid. Rejecting cookies only helps so much. Which service shares what, and how to clean up shared chats, cookies and your Android advertising ID.

https://www.notebookcheck.net/ChatGPT-Claude-Grok-Your-chats-end-up-with-trackers.1421879.0.html

[Running AI locally]

Don't want to be tracked or whatever? Did you know that you can run AI locally (and privately)?

Ideally all you need minimally (it even rhymes, so it must be true), is a gaming laptop with 32 GB RAM and a 6-8 GB VRAM GPU and it can then fit a lot of context, too:
Calculating of available context tokens of existing frontier models (Qwen, as of this writing)[1], using an approx. minimum recommended Q4_K_M quants on 32 GB RAM+8 GB VRAM devices (16,384 [KV cache context tokens per GB[2]]*(32 [GB RAM] + 8 [GB VRAM] - 7 [GB for the OS[3]] - [GB quant filesize])):

• For Qwen3.6-35B-A3B (MoE) (eg huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF):
  178,500 context tokens = 16,384*(32+8-7-22.1)

• For Qwen3.8-27B (dense) (eg huggingface.co/unsloth/Qwen3.8-27B-GGUF):
  270,000 context tokens = 16,384*(32+8-7-16.5).

(If Qwen4 gets a reduction in KV cache context tokens memory requirements, like in [4], then the calculated context number will further increase.)

[1] artificialanalysis.ai/models/open-source/small
[2] reddit.com/r/Qwen_AI/comments/1vo8pjz/qwen3827b_kv_cache_works_out_to_64_kibtoken_so/
[3] techpowerup.com/353395/windows-11-26h2-update-lowers-ram-consumption
[4] huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash (see "Global KV Cache Per Token (Bytes)" image)

(due to anti-spam system, 2nd part is in next post)

[Running AI locally]

And these ~30B parameters models are no joke:

Once again anthropic doing free marketing for open weight models

Models of 30B parameters and under can already complete many of the individual steps for replication today, though they make mistakes when composing them end to end. In our simplified evaluations, GLM-5.2 and stronger models are capable of end to end replication. If small models improve at their historical rate, they'll match this capability level by the end of this year. Small weight-based replicators may be possible soon after.

And the top rated comment from that reddit post is:

they talking about our goat 27b

Communities e.g.:
reddit.com/r/LocalLLAMA/top
reddit.com/r/LocalLLM/top

More about running AI models locally: notebookchat.com/index.php?topic=315954.0 ("Your own ChatGPT, offline: AI without the cloud on your laptop") and comments.