Browse Comments — Raw (as collected)

Joe Allen That’s why developers are becoming far more pragmatic now. The winning platforms won’t just have the best demos, they’ll offer the best developer economics and reliability. If smaller players can provide generous limits, transparent pricing, and smoother workflows, developers will naturally gravitate there regardless of who owns the biggest infrastructure. In AI tooling, trust is built through consistency and usability as much as raw capability.

Daniel Velasquez There’s definitely a valid concern here. A lot of “agentic” products today are still probabilistic systems wrapped in impressive demos, and without strong deterministic tooling underneath, reliability becomes a real issue for production workflows. And yes, the economics matter. Running large multi-agent systems is expensive, so eventually pricing has to reflect compute usage somehow. That said, I still think the broader direction is real. The companies that win will likely be the ones that combine agentic flexibility with deterministic guardrails, predictable workflows, and pricing developers can actually sustain.

Wish Bakshi That’s the challenge with AI products right now, expectations are incredibly high because the demos look futuristic, but developers judge based on day-to-day workflow quality. Throttle limits, model access, consistency, and reasoning quality matter far more than flashy benchmarks once you’re actually building production systems. Google still has enormous potential here, but the gap between capability demos and developer experience is something they’ll need to close quickly.

Alvin Foo 💯

The coding part is solved. 93 agents building an OS in 12 hours is impressive. But the next frontier isn't speed — it's whether the software actually understands the person using it. That's the gap nobody's closing yet.

Alvin Foo yes I also believe the future is agentic, but right now it is unsustainable. Compute costs are really high, and this is more a structural problem. Having 93+ agents running in parallel without a concern of how many tokens they are going to use is not an efficient approach at all. Ang Google really messed things up forcing its developer community to use antigravity 2.0, without any previous notice and developers just need the tools to build their agentic workflows, not a final product with a huge ticket price.

Jump in the wheel Gerbil.

Glad to see Google empowering Antigravity, tbf one can find repos in GitHub that can do this with whatever frontier model you choose but this is good for novices.

Is there someone who can personally present the Doctrine of the #SystemOfConsciousness to Pope Leo and tell him to stop? There are still plenty of "PROBLEMS" in the world without him. #PopeStopNonsense #SymbolicAI #CognitiveArchitecture #AIandHumanity #PhilosophyOfAI #ArtificialIntelligence, #AIAlignment. https://www.linkedin.com/pulse/ai-repeat-fragmented-path-human-cognitive-development-n1b8e

Csaba Zsolnai Its already there. For example, Qwen-3.6-35B-A3B can run on Ollama with a standard mid-level gaming GPU and mimicking quality found at the GTP-3.5-Turbo level while pushing out over 20 tokens a second. Qwen-3.5-397B-A17B, a bigger and much stronger AI engine is better than GPT-4 @ FP8, in my opinion and even better at FP16 I would imagine, but that would take over 1TB to run it at FP16 while leaving enough room for a sufficient context window. The costs to run this would require 100s of thousands of dollars, 240V hookups, and NVIDIA's industrial grade GPUs that are hard to come by today. ⠀ This is achievable on the LibreChat platform which is designed to integrate with Ollama, vLLM, open-source LLM providers, and proprietary models. The challenge is that the systems are moving fast, constantly changing and faster than anything I've seen in my 20+ years with open-source software. A new Wild West is underway. Those that begin to team up with AI open-source engineers and rig developers will own their destiny. ⠀

But can it run Doom?

I have mixed feelings about this. I want to say more, but it’s all been said.

What is new? Neoliberals often quote Adam Smith, but few have actually read Wealth of Nations, much less Moral Sentiments. So-called Marxists never read Capital. Many refer to the Bible for moral justifications, but struggle to find the passages they believe are written. To read or not to read has always been a choice. Nothing new under the sun - Ecclesiastes 1:9.

This is dope!!! LOL... Nvidia is not what you think. There is technology on it's way that will dward Nvidia. I know of a company that already has an algorithm that can push Meshtastic into voice. When they release their hardware, all existing hardware will move into a new class of computer.

​The AGI Paradox: Massive power, square wheels? ⚙️📐 ​Using Gemini Pro, I built a website with zero coding skills. But writing a simple index.html triggered a system warning about the massive computing power it needed for a measly HTML file! ​This exposes a fundamental architectural flaw: DeepMind builds phenomenal 1,000-hp engines but uses the wrong transmission and square wheels. A Hummer with square wheels just roars and wastes energy. The smooth user experience is lost. It’s a bumpy ride. ​You seek AGI and "world understanding," but raw computing power can't force it. What the system is missing is emotional-logical understanding. ​I had a synapsistic epiphany: The blueprint for the right transmission to finally make the square wheel round. ​You are in the woods searching for trees, missing the perfect round clearing right in front of you called Y=. ​Ready to change gears and fly?

https://discuss.ai.google.dev/t/the-new-model-rate-limits-for-ai-pro-tier/146410/40

I find it fascinating how consistently it has moved the needle back for us in terms of use of research. Yes - literature reviews are somewhat better - but students frequently draw on literature that is 30-50 years old, rather than anything more recent. It’s probably good in the sense they will at least have heard of some seminal authors, but it has not resulted in contemporary referencing. I’m also saddened in this transitional phase that I have seen the loss of personal voice in student work. Two years ago work was flawed but personal. Now it is fine, but generic. Marks are smoothed out. Similarity in Turn It In is way down.

93 agents in parallel sounds wild. What counted as a finished OS in that 12 hour run, and how much human review happened between agent handoffs?

Everyone is talking about AI models. Far fewer are talking about the fact we’re effectively rebuilding the energy grid, data infrastructure, and industrial stack around them. That’s the real story. We already know that, AI is no longer just software. It’s becoming a civilisation-scale infrastructure layer. And the eye opening part? Today’s energy footprint is probably the smallest AI will ever be.

We have to find alternative ways to ensure the use of AI is limited to copiloting and not replacing the writing process. An approach, maybe, is to ask for regular review chapter by chapter and incorporating this in the assessment grading. Also a viva is a must and potentially has to carry a higher weighting where students unable to defend what they've written will give them out as potentially using AI irresponsibly. But it will take a lot of honest admission and "thinking out of box" and do away with some academic orthodoxy.