Google announces Gemma 4 open AI models, switches to Apache 2.0 license
I do not really like to take X posts as a source, but it's Jeff Dean, maybe there will be more surprises other than what we just got. Thanks, Google! Edit: Seems like Jeff deleted the mention of 124B. Maybe it's because it exceeded Gemini 3 Flash-Lite on benchmark?
Gemma 4 has been released
[https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) [https://huggingface.co/unsloth/gemma-4-31B-it-GGUF](https://huggingface.co/unsloth/gemma-4-31B-it-GGUF) [https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF](https://huggingface.co/unsl
Gemma 4 and Qwen3.5 on shared benchmarks
Anthropic Says That Claude Contains Its Own Kind of Emotions
and what does this even mean? "internal representations of emotion concepts driving claude behaviour" I get it that they don’t feel emotions and they simulate patterns of emotion, but the scary part is humans respond to the simulation the same way "panic"
OpenAI acquires TBPN, the buzzy founder-led business talk show
OpenAI acquires TBPN to accelerate global conversations around AI and support independent media, expanding dialogue with builders, businesses, and the broader tech community.
Microsoft Readies $10B AI Investment Plan in Japan
Lemonade by AMD: a fast and open source local LLM server using GPU and NPU
171 emotion vectors found inside Claude. Not metaphors. Actual neuron activation patterns steering behavior.
https://preview.redd.it/kkvvcqr8susg1.jpg?width=1200&format=pjpg&auto=webp&s=ae0315c528afef84c035354927c4b9c5d8ec0bb4 Anthropic's mechanistic interpretability team just published something that deserves way more attention than its getting. They identified 171 distinct emotion-like
Why OpenAI’s Fidji Simo Bought the TBPN Podcast Amid Crusade Against ‘Side Quests’
OpenAI has purchased TBPN, an online talk show that often interviews AI executives and other tech leaders. The show goes live every weekday at 2PM PT, often for a three-hour duration, counting OpenAI CEO Sam Altman, as well as executives from Meta, Microsoft, Palantir, and Andreessen Horowitz, among
AI-2027 forecasters move their timelines ~1.5 years earlier, predict 2027 or 2028 most likely year for AGI
Link to twitter thread: https://x.com/eli\_lifland/status/2039773600555979251?s=20 Link to blog: https://blog.aifutures.org/p/q1-2026-timelines-update
'Backrooms' and the Rise of the Institutional Gothic
One of the best sensible reasons that I can think of to have an llm downloaded on my cell phone would be emergency advice.
It seems like every conversation about derestricted models everyone treat you like a pervert. The fact is you can be sensible and be a pervert 😂.
A $20/month user costs OpenAI $65 in compute. AI video is a money furnace
New laws to make it easier to cancel subscriptions and get refunds
Gemma 4 is efficient with thinking tokens, but it will also happily reason for 10+ minutes if you prompt it to do so.
Tested both 26b and 31b in AI Studio. The task I asked of it was to crack a cypher. The top closed source models can crack this cypher at max thinking parameters, and Kimi 2.5 Thinking and Deepseek 3.2 are the only open source models to crack the cypher without tool use. (Of course, with the closed
Altman on shutting down Sora: 'I did not expect 3 or 6 months ago to be at this point we're at now; where something very big and important is about to happen again with this next generation of models and the agents they can power.'
[https://youtu.be/mJSnn0GZmls](https://youtu.be/mJSnn0GZmls) ‘We have a few times in our history realized something really important is working, or about to work so well, that we have to stop a bunch of other projects. In fact, this was the original thing that happened with GPT3. We had a whole por
Gemma 4 running on Raspberry Pi5
To be specific: RP5 8GB with SSD (but the speed is the same on the non-ssd one), running [Potato OS](https://github.com/slomin/potato-os) with latest llama.cpp branch compiled. This is Gemma 4 e2b, the Unsloth variety.
Gemma 4 26b a4b - MacBook Pro M5 MAX. Averaging around 81tok/sec
Pretty fast! Uses around 114watts at its peak, short bursts as the response is usually pretty fast.
My first impression after testing Gemma 4 against Qwen 3.5
​ I have been doing some early comparisons between Gemma 4 and Qwen 3.5, including a frontend generation task and a broader look at the benchmark picture. My overall impression is that Gemma 4 is good. It feels clearly improved and the frontend results were actually solid. The model can
Gemma 4 is seriously broken when using Unsloth and llama.cpp
Hi! Just checking, am I the only one who has serious issues with Gemma 4 locally? I've played around with Gemma 4 using Unsloth quants on llama.cpp, and it's seriously broken. I'm using the latest changes from llama.cpp, along with the reccomended temperature, top-p and top-k. Giving it an article