NVIDIA just released an optimized GLM-5.2 on Hugging Face
A 753B parameter MoE with 1M context,
quantized to NVFP4 for Blackwell GPUs—
nearly matching FP8 accuracy.
The AI research & tweets I’ve been sharing always pointed to a broader shift - the dominos.
Anywho ignoring means u also missed $SNDK, $silver, $nvda , $WDC .
Well, ree - don’t underestimating storage, compute & data in what’s coming next. dyor, I hate ticker posting.
Gl Hf
Cuz it’s art, that comes with livid emotional experience.
i get most frustrated with it as i’m autist & i found “writing boring” & thought Ai will fix that but it hasn’t xD ig cuz Experience isn’t same as learning. It’s the subtle gap between felt understanding imo. :D the “why & how” things are presented. Just sucks balls ngl lmfao.
We're rolling out Brain: a self-improving context-graph of all your sessions, connectors, and files.
Brain updates itself overnight with fresh context proactively, and feeds itself to every task on Computer, allowing Computer to be stateful and self-improving.
Available to all Perplexity Max subscribers.
I strongly agree with this mostly. Humanoids only make sense because of “familiarity bias”.
i’m bearish aw over a simple fact that they’re counterintuitive. 🤷🏽♂️ Unless are we really expecting every other machine to stay unintelligent, human-controlled, or dependent on 2 hands & legs?
Here’s a counter ; image 2-3-4 are better at majority of work with bolted on attachments.
Robotics // physical Ai is at heat but focus on Humanoids is misplaced bet, i love too cuz grew up watching em in movies.
Verticals will cluster around highest-ROI applications & usecase ones, not one universal body. Use-case specificity is what actually decides. There simply aren’t enough serious humanoid platforms yet to be properly bearish on the category yet but yes I absolutely dislike the fomo. At the same time, not every robotics company should pile into building the exact same two arms and two legs.
The real question isn’t “can we make everything human-shaped?” It’s “what embodiment delivers the highest ROI for a real use case? - will the user even be able to pay for everything extra - should they?”
Humanoid is a form factor, not the market. A fork is better than a spoon for some jobs & vice versa. Mobile phones & desktops have shown us a good dataset here.
Majority are better served by task-specific machines that just do the job better and cheaper. Some industries & some of us need humanoids.
I’ve been a firm believer of Applied intelligence & machines will reshape. Are 2 hands & 2 legs best shape for specific tasks? How about 4 arms 1 brain for handling aviation instead of 2 humanoids?
I believe people will eventually assemble whatever they need themselves & the winning companies will be the ones making the software that makes all of it work better & that’s pretty the core thesis of doing what we do.
I Don’t get the maxi’s of any specific shape. It’s not expected for arms to be able to walk because they can perform surgery better.
Yes, a lot of people will want generic solutions. But the highest usecase products will almost always be specific.
Humanoids are kind of a counter argument for intelligent first machines & also because human shape isn’t perfect for robots at scale. ( call me a shapist 🤣 idc )
if you take humanoid maxis at their word, they’re the swiss army knife of robots
but if they can be easily slotted into any role, their deployments have no stickiness bc a competitor can also easily replace them
they have to monopoly win or margins go to zero
bearish
Oh My My, @satyanadella just gave a green flag blasting it out to everyone why this matters a lot!
Everyone will converge! As accelerants, what a great piece of writing about ecosystem.
The reason of everything written above is well written here -
https://t.co/t38uJUhCWZ
Ai labs, tech giants & researchers keep converging on the same conclusions independently. Certain directions in AI seem to exert their own gravity - take different paths, but end up in remarkably similar places. It’s an uncharted territory Every now and then I stumble across an old tweet and realize I’d been talking about the same thing months ago & Vice versa
Nice to finally have a name for something sitting in our R&D logs since Nov 2024 -
Fable-5 level outcomes with Fusion as OpenRouter calls it ( Concept of Multi LLM harness for better output)
Btw “Harness” term didn’t exist publicly until few months later.
Concept we discussed at @aiDotEngineer Jan 2025 [ https://t.co/OqHDaWvp8r ] Guess real building and frontier exploration are still massively -EV relative to brainrot retardation content personally. Probably an algorithm discovery problem more than anything, so hard to blame @nikitabier, but Doesn’t make it any less painful to watch though & needs fix
Attending the @aiDotEngineer summit back in February was a huge boost for @ProbeAgi ( for building @agiObjective ) . The connections, the deep tech conversations, and being at the epicenter of top AI giants and the brightest minds directly shaped our direction and strengthened our resolve to build ProbeAgi and Objective into what they are becoming.
Speakers this time specially Jensen Huang (@nvidia) has been amazing. I missed this one live, but I will not miss the next. Agents, swarms, and vast frontiers of technology remain untouched. We are only beginning to see their true power. While we advance toward generative AI, the real question is how institutions and individuals can use it better. That challenge is still a vast and open playground.
World is prepping for better usecase for Ai now! Future towards AGI is not just a better chatbot . It is a thinking engine that can reason, learn, and act across any field with human-level or greater skill. It can solve any problem, adapt in real time, and create solutions without being locked into one role.
AI answers often match the quality of our questions, but real progress is when they go beyond, reading context, sensing nuance, and choosing the best path forward. That is the doorway to the future of AI and a step toward Artificial General Intelligence. All LLMs now seem to agree on this. Laughs, yes, it was once all about self-promotion, but those days are behind us as the community becomes more aware.
Now imagine a hive of leading LLMs @xAI@OpenAI@AnthropicAI and more in constant feedback loops, refining each other’s answers and converging on the optimal result for every challenge. This is @AgiObjective, a collaborative intelligence engine where efficiency and outcomes matter most.
Nice to finally have a name for something sitting in our R&D logs since Nov 2024 -
Fable-5 level outcomes with Fusion as OpenRouter calls it ( Concept of Multi LLM harness for better output)
Btw “Harness” term didn’t exist publicly until few months later.
More & where we showed it first - How we use it for Deep research / Harness & building internal products & learning at @OnzoreAI / Multi Agent Orchestration & outcomes & much more -
https://t.co/GxGMbWAzZH
https://t.co/RhAO8r9VkJ
https://t.co/1lh2qBSO3y
A few thoughts is where observers start; experience is where informed opinions begin.
Alignment’ll be easy to explain as screenshots from 2024 are what being discussed about “agents” & moltbook & this topic is drifting into the exact failure mode @balajis is describing.
It’s not a new experiment & it’s not bad but what’s the outcome?
The perspective I am sharing comes from a team with over half a decade of experience engineering data and AI infrastructure for fintech giants and Fortune 500 environments. At that level of scale, the “nurturing” narrative breaks down quickly. In high stakes production, systems are not nurtured; they are architected to survive failure modes that most discourse has not yet encountered. Screenshots are attached for reference.
Treating “nurturing mixtures of experts” as a novel frontier misses the industrial reality. We explored and stress tested these patterns, and by October 2024 we had already moved past this phase. Progress required coordinated team execution, aligning on orchestration, verification, and failure handling rather than conversation. When the objective is bottleneck removal, orchestration is the primary lever. This is not theoretical.
Much of the current commentary reflects a “monkeys on typewriters” dynamic ( as @hosseeb said ) , high volume, low signal loops driven by observers reacting to prompts rather than builders reacting to years of deployment data. The exchange under @frankdegods’ comment is representative, reactive interpretation layered on reactive interpretation, without ownership of orchestration or production constraints. Ease of access to information has reduced the cost of commentary, not the cost of execution. Engagement and FOMO are not substitutes for quantitative output.
The reality from the field is straightforward:
• Agents are downstream of prompts.
• Unmanaged interaction drains tokens, time, and capital.
• The hard problem is not conversation; it is orchestration.
Treating observer level noise as a breakthrough obscures the engineering required to make these systems viable. @balajis is correct. Without explicit coordination, you are observing agents hallucinate in an expensive loop. These experiments are fine as exploration, but the industry moved on months ago. What matters next is not reacting to information, but advancing the orchestration layer itself.
The screenshots linked below are from @probeAgi’s internal work from October 2024 and are unedited. The focus now has changed & razor sharp on, intent-to-execution at the orchestration layer & defensibility ultimately comes from moats, momentum & experience + since we’re planning to go beyond,
Happy to discuss the architecture and what we are building next in depth in dms . I’ll refrain from detailing the specifics in the open & i had realized that it’s beyond most people’s attention & understanding here, and they only tend to realize it months later... like clockwork. ^^
https://t.co/c5jLmMAmid