![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
LOL!
Google's Gemma 4 LLMs are pretty friggin sweet. They are holding their own for daily driver models for Agent Computers. Have not ever used any of their normie stuff.
Right? I wanna ambush him tomorrow and experiment, like "hey Google, play any song from Big Black's magnum opus album Songs About F*cking" to see if the stupidity is deeper than clbuttic surface level stuff.
These are all pre-trained, right? I've got a 16GB RX 6800 I could play with and I hear that Google just release a Gemma 4 12B that will run on 16 GB.
Gemma 4 12B came out in June. You can run the 8-bit in 16GB for sure. Gemma 4 12B dropped yesterday. (Encoder-Free Advantage (unique to 12B in the family):No heavy separate vision/audio encoders. Raw image patches and audio waveforms go straight into the LLM via tiny linear projections.→ Lower latency, lower memory overhead for multimodal tasks, and much easier fine-tuning (one model, all modalities updated together).) The scuttlebutt is that this one is as good as 26B if you don't need the AV encoding.
The issue is everybody developing AI are just viewing it as a product to sell, which is why it will be a long time, if ever, there will be actual useful things it can do. This combined with the hard coded guardrails in LLM, agentic and generative AI models means AGI will never be a reality, at least in my lifetime.
Gemini is not the worse AI, google tend to often have the best AI for many things (from folding protein, predict weather,
The really not how Google has seen it... the AI that has been helping google search or youtube suggestion or gmail autocomplete for a decade were always seen as a tool directly made for something useful, same for AlphaFold (that they did for free).
My investment position is simple.
From a business standpoint, I don't think it will be useful for every task, but it can help augment a smaller workforce to accomplish better quality or industry results. We have already seen it with Expedition 33. The game was only controversial to the industry, not because the quality was subpar, but because it was a AAA game made by a small studio. That is scary stuff for the industry. And if Expedition 33 can find success because of AI augmentation, casting AI as a small business failure is not a reflection of reality. I don't even think we have seen the tip of what is coming. Local LLMs are just now getting an exploratory phase by the people that will build on top of them. We're just getting started.
very short term maybe, usually tech improvement that everyone can access does not make things more profitable.... freezer did not made grocery chain more profitable despite the giant gains... everyone has them. We can line up all the great tech advancement in grocery stores supply chain over the last 200 years, they still are ran at 1-3% profit margin. The idea that AI will be seen in profits growth or lack of them tend to be a misunderstanding of tech advancement and how it create wealth or in general of what competition do. If we are lucky it will make them in general less profitable even, because of how easy competition and entering market will be.
Pricing in general are extremely sane/prudent in that space, Nvdia, Microsoft, Micron, TSMC, google, meta, broadcom, etc.... are all cheap relative to forward operating income.
Forward is often from sales already made (people book early), but even the price relative to the old operating income is quite low in that space for growing company (AMD excluded), which make sense, expecting margin to fall a lot do make sense. and i am comparing to what were also speculation at the time (when it come to stock price... of course).
of course ? as they are future affair, stock price is speculative by nature.
If you're making the argument that this isn't bijective to the dot com bust, then we're arguing different points here.
We all agree that is not at all similar to the dot com bust in term of stock valuation of the big players.
It's pretty good. Google's is one of the more affordable options compared to Claude or OpenAI (and their free tier is quite generous), especially if you use it in something like OpenCode or Hermes, which are both very token efficient. My bashing of the Google models was Gemini ... the magical hallucinating garbage chat experience ... not coding. But I've been trying it with agentic apps and it's much better in that regard.