Recent searches

No recent searches

Search options

Only available when logged in.
101010.pl is one of the many independent Mastodon servers you can use to participate in the fediverse.
101010.pl czyli najstarszy polski serwer Mastodon. Posiadamy wpisy do 2048 znaków.

Administered by:

Server stats:

474
active users

101010.pl: AboutStatusProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.5.0-alpha.2

#ollama

7 posts7 participants0 posts today
Denzil Ferreira

Played yesterday with gpt-oss:20b on my home lab. Ollama by default uses 2K tokens as context so those are used up quite quickly since this is a reasoning model. Increased to 128K (enough to fit a whole book) but then it would take forever to start producing an answer. Setting it to 64K was a sweet spot, and now it's quick, connected online and has vision understanding. To test, added a pdf and asked questions about it's content and it worked flawlessly. Using 9070 16GB on a 32GB RAM machine.

Using Continue on VSCode, I keep my AI assistant in-house. No subscriptions, all here.

#homelab#gpt#ollama
Replied in thread
*
Sepia Fan

@kjhealy

Locally tested with #Ollama in German with #Gemma3 (Google LLM) for "Blaubeere".

✅️ Wrong letter count
✅️ Wrong letter positions
(Pic 1)

But if forced to count via "list all letters and then tell the count of X" the #LLM seems to be able to report the correct answer. (Pic 2, two restarted instances)

C++ Wage Slave

If you think MP3 sounds good, choose a song you love that has a detailed, spacious sound, and encode it in #MP3 at low bandwidth. Hear the jangly tuning, the compression artifacts, the lack of detail and stability and the claustrophobic sound. Now that you know it's there, you'll detect it even in MP3 samples at higher bitrates.

This toot is actually about #GenerativeAI. If you can, download #Ollama and try some small models with no more than, say, 4bn parameters. Ask detailed questions about subjects you understand in depth. Watch the models hallucinate, miss the point, make logical errors and give bad advice. See them get hung up on one specific word and launch off at a tangent. Notice how the tone is always the same, whether they're talking sense or not.

Once you've seen the problems with small models, you'll spot them even in much larger models. You'll be inoculated against the idea that #LLMs are intelligent, conscious or trustworthy. That, today, is an important life skill.

Teodor Gross

I invented `.awesome-ai.md` - the new standard for AI tool discovery!

Like `.gitignore` for Git, but for AI tools on GitHub . My system automatically scans the entire GitHub and discovers new AI projects.

Website: awesome-ai.io/submit-info
Repository: github.com/teodorgross/awesome

Awesome AISubmit Your AI Tool | Awesome AISubmit your AI tool via GitHub integration or website integration with hidden div.
#AI#GitHub#Innovation
Markus Eisele

Ollama v0.10.0 is here! Major highlights:

- New native app for macOS & Windows
- 2-3x performance boost for Gemma3 models
- 10-30% faster multi-GPU performance
- Fixed tool calling issues with #Granite3.3 & Mistral-Nemo
- `ollama ps` now shows context length
- WebP image support in OpenAI API

github.com/ollama/ollama/relea

GitHubRelease v0.10.0 · ollama/ollamaOllama's new app Ollama's new app is available for macOS and Windows: Download Ollama What's Changed ollama ps will now show the context length of loaded models Improved performance in gemma3n mo...
#Ollama#AI#LocalLLM
TrendingLive feeds
About