Chief Agenteer

Later this year, or early next year people will say that they got their AI to beat GTA 6 in under five hours so that must mean we've reached AGI. It won't be true but the first part of what I said, a delayed GTA 6 launch, is far more likely than AGI.

The Hype vs. The Reality of J-Space: ​Anthropic’s J-Space discovery is getting a lot more credit than it deserves. The reality is that this internal "workspace" is entirely dependent on the prompt making sense. It isn't an independent thinking realm; it only forms a coherent structure when you feed the model structured, predictable language. ​The "Gibberish" Test: ​If you give the model a normal, logical prompt, J-Space works exactly as advertised, holding onto math steps, hidden plans, or context. But what if you feed it incoherent gibberish? The whole mechanism collapses into noise. ​Unlike a real mind, the model doesn't get "confused" and try to puzzle things out. The initial layers process the raw data, but the J-Space workspace just turns up empty or fragmented. Because the system's only goal is to predict the next word based on patterns, a lack of patterns means there is simply no "thought" worth broadcasting. This isn't a glitch; it is exactly how a statistical pattern-matching engine is supposed to act. ​Why It Is Nothing Like a Human Brain: ​This total reliance on context highlights a fundamental difference between AI and human cognition. A human's "global workspace" doesn't shut down just because the input is meaningless or ambiguous. ​We process confusion: If we hear noise or see something we don't understand, our brains generate error signals and actively try to make sense of the ambiguity. ​We don't just use words: We rely on visual imagery, spatial reasoning, and emotional processing, none of which require language to function. ​We have background systems: Massive unconscious systems (like memory consolidation and physical regulation) continue to operate completely independent of whether our conscious thoughts make logical sense. ​The Bottom Line: ​J-Space has no fallback. It has no deeper, non-linguistic processing to lean on. When it can't deduce the context, the workspace simply blanks out. ​Ultimately, this proves that LLMs do not operate like human brains. They are incredibly sophisticated statistical engines tied completely to the coherence of their training data. The fact that an internal workspace emerges under those strict conditions isn't proof of intelligent cognition. It is just what happens when you mathematically train a machine to guess the next word really, really well.

The "loop" has existed in the SDLC since the SDLC was invented. Nothing to do with AI or who does the coding. Coding, testing, refining against a functional specification and an architectural framework is what professional developers have done forever. This "loop" stuff on X the last few days is just clickbait for amateurs. In my own situation, I don't ever run prompt to prompt. Everything follows SDLC. If you follow the SDLC you don't need social media to learn stuff you already should know as a developer. The only difference is how people set it up with AI and how humans do the work manually. My method is to have the harness process a methodology file at each turn during the development process. That markdown forces the agent to review requirements versus test results and keep going until the functional specification is fully delivered. You don't need separate programs to define loops. Just a methodology markdown. Loops happen on their own if you know how to setup the development methodology correctly.

⚠️Pretty amusing that the tech executive list traveling to China doesn't include anyone from ASML. The real story is that the US is trying hard to sell specific tech into China but at the same time, preventing access to technology produced by ASML, which produces the tech that produces chips. ASML is the key player when it comes to making AI chips at scale. Just like what happened with the NVIDIA trade ban, this strategy will fail. When NVIDIA was banned from selling certain products into China, Chinese companies like Huawei doubled down on their own stack and on performance can now compete with NVIDIA. The gap is in scaling production and that's where ASML fits in. The problem for the US is that by banning ASML exports to China, China will double down on the production of homegrown technology to perform the same tasks that ASML machines perform. This is already in progress. It's quite likely that within 18 months, that China will match or even beat the capabilities of ASML machines. It's a damned if you do and a damned if you don't situation for the US. Allow access to ASML and the Chinese will build homegrown capabilities even faster. Continue to disallow access and the Chinese will still catch up but a touch slower. What's clear here is that given the depth of engineering skills now native to China, that it's a foregone conclusion that eventually the Chinese will overtake the technology leadership in both AI and chip production just like they have beaten the US in almost every other industrial segment this far. And there's nothing anyone can do about that.

🆘 Consciousness doesn't operate in "turns". Think about that.

I'm glad I'm not one of the idiots who listened to the idiots who told everyone to go and buy Macs. I bought Blackwell instead. Never trust anything you read on X blindly.

🚨 If you want to retire rich before the age of 45 you don't need 𝕏.

If you think 𝕏 is educational, you're going to find out the hard way that you learned all the wrong things here.

Big cloud providers are building excessive capacity that has a high probability of being underutilized and then hardware obsolete before any ROI is achieved.

As an Australian I think the government here needs to go further and suggest to the US that any further financial damage to the citizens of Australia through US foreign policy would mean a rethink of the alliance entirely. Quite honestly Australia is better off having a security pact with China, who is our most important trading partner.

Dude I've got models running locally on 8GB RTX 4060, 16GB RTX 5060Ti, RTX PRO Blackwell 4000 24gb and DGX Sparks. Some of the PC hardware hosting the cards goes back to 2014. If you want to just play around grab the 16gb RTX 5060Ti and play with the smaller Qwen and Nemotron models. I've got AI based OCR business solutions running on 8GB VRAM cards for customers. Keeps their data off the cloud and private.