The absolute state of AI results.

[Ekco]

the machine god weighs in

Breaking it down by what's actually checkable vs. what's a real unresolved argument:
Ekco's infrastructure claims hold up. All three check out against current reporting:

Data centers: about 12 gigawatts of data center capacity is expected to come online in the U.S. in 2026, but only about one-third of that capacity is currently under active construction, with analysts estimating 30-50% of planned 2026 capacity will be delayed or canceled, mostly due to difficulty sourcing electrical equipment and securing power. "Only half started" is actually generous — it's closer to a third.

Memphis: confirmed, repeatedly. xAI installed 35 gas turbines capable of producing 420 megawatts alongside Tesla Megapack battery systems because the site's initial 150MW capacity was covered by only an 8MW grid connection. This has also drawn lawsuits and air-quality complaints from the surrounding community.

China nuclear: roughly right. China operates 58 reactors and has 33 additional units under construction as of early 2026, with other counts running as high as 60 operating and 36 under construction. "60 and 40" is a slight round-up but in the right neighborhood, and the trajectory (10x growth in two decades, aiming for 110GWe by 2030) supports his point about pipeline depth.

The "10-12 million power users" figure is plausible as a strict definition but it's worth flagging that headline AI usage numbers are way higher (179 million Americans have used some AI tool) — the 10-12M figure only makes sense if it's measuring something narrower like daily, deeply-integrated professional use rather than "tried ChatGPT once." Without seeing that source's exact methodology I can't grade it, but it's not an absurd number for that narrower bar.

The scaling/plateau argument is where it gets genuinely contested, and both of you are half right.

Ekco's right that naive dense-parameter scaling has hit diminishing returns — this isn't a fringe take, it's the documented reason labs pivoted to MoE, reasoning/test-time compute, and synthetic data. In 2021, doubling compute reliably doubled measurable capability on most benchmarks; by 2025, doubling pre-training compute buys maybe a 10-20% improvement on the hardest reasoning tasks. That part of his argument is mainstream, not contrarian.

But BradZax's "progress increased, not flattened" also has real data behind it — just on a different axis. METR's tracking shows no evidence of the exponential growth in AI task-completion time horizon slowing down, and if anything the trend accelerated in 2024-2025, with time horizons doubling every 4 months instead of every 7. So overall capability (measured by what models can autonomously complete) hasn't plateaued — it's the naive scaling method that plateaued, while labs found other axes to keep climbing.

On the "models bunching at the top of the curve" chart specifically: that's a real, debated phenomenon, but it cuts both ways. It could mean genuine convergence toward a capability ceiling. It could also just mean the benchmarks themselves are saturating (hitting 95%+ scores) while real-world capability gaps are still widening underneath — METR explicitly flags this as a live possibility, that benchmarks could be saturated without leading to AGI, which is closer to your read, or that the fast pace simply continues. Nobody's resolved which is happening.

The "AGI/ASI via transformers is sci-fi bullshit" claim is the one place I won't hand a winner — that's a live, roughly 50/50 split among serious researchers (LeCun-style skeptics who think you need a fundamentally different architecture vs. people who think scaling + RL + tool use gets there), not something settled by the current data either way.

[Duik]

Now that the (largely) freebie tokens will dry up because billionaires want they money back those of you that ask A.I agents stupid questions like Summarize the posts of forum user <insert user> and give me bullet points as to why they are a boomer for free will defo not wanna pay for that shit.

A.I agents need to make money for their digital pimps.
Start paying.
Medical imaging manipulation/interogation? Good usecase.
Vibe coding with guided prompts by knowledgeable programers? As much as i hate to admit it. Good usecase.

Putting a suit and tie on ya cat photo? Go for it, but be prepared to pay for it with little to no returns.

Welcome to A.I.

[Ekco]

that graph now that ive actually look at it is just about time on task and stops at 16 hours dues to unreliable test suite, whatever the fuck that means, write a better test, either way its outdated as fuck Claude runs for days now

Claude can run continuously for days on massive coding projects thanks to dynamic, agentic harnesses like the Claude Agent SDK or Claude Code

so whoever put that chart together doesn't know how people have been using the models for numerous months now, in parallel with a overseer agent and dozens of sub agents and all that bullshit in an agentic self-correction loop

the thing that actually matters is capability and the plateau is way more pronounced in those charts in the models released in the last year, the big gains are in reducing the time to complete a task successfully, some giant codebase can take chatgpt 5.5 3 days to work on and Mythos supposedly did the work correctly in like 10 hours or something

[You must be logged in to view images. Log in or Register.]
we've hit that ceiling, the is AGI even possible easily by just making a 1 trillion parameter model type idea and the answer is no and charts like this are pointless now because of diminishing returns of just training larger and larger parameter single models Qwen opensource is at like 350b parameters but that's probably just adding up all the separate MoE models

While frontier closed-source providers (Anthropic, OpenAI, Google) no longer disclose exact parameter counts for models like Claude Fable 5, Claude Opus 4.8, or GPT-5.5 Pro, the landscape for open-weights and verifiable models has scaled significantly heading into mid-2026.

Frontier Open MoE Qwen3 235B A22B / Qwen3.5-397B-A17B 235B – 397B total (17B–22B active per token)

yeah, they stopped reporting the parameters because number no longer going up = scary for investors interested in two companies about to IPO

Major Paradigm Shifts Since Late 2024:

Active vs. Total Parameters (MoE Dominance): Large open-weights models have shifted aggressively toward Mixture-of-Experts (MoE) architectures (seen in the Qwen3 and DeepSeek-V4 series). A model may have up to 397B total parameters sitting in storage, but only routes ~17B to 22B active parameters per token, drastically reducing inference latency while preserving massive knowledge depth.

The "Medium" Sweet Spot Shift: The traditional 7B baseline has moved up. Architectures like Gemma 3 (12B) and Qwen3.5 (9B) maximize dense compute efficiency, effectively rendering the old 3B–7B performance tier obsolete for complex multistep coding or agentic loops.

so the sweet spot, is the same model im using for Kaia on a GPU from 2021 that costs like 250 bucks, a model running LOCALLY on your cell phone has enough juice for 99.9% of user queries if built right to use tools like let_me_fucking_google_that_for_you.py considering what most people are actually uses these chatbots for

so not only is AGI/ASI not going to happen, all these companies are going to go bankrupt causing a deep recession because their business plan we started this journey with doesn't make any sense anymore

open source models wins on both ends of the spectrum, locally run open source wins for a non trivial chunk of the consumer/enthusiast market and enterprise coding just got something that costs 1/5th of a Claude or ChatGPT seat dropped in their lap thanks to China

[BradZax]

Originally Posted by Ekco [You must be logged in to view images. Log in or Register.]

so whoever put that chart together doesn't know how people have been using the models

The people that put the chart together are the ones that made the models that are throttled 99% those people have been using.

[Ekco]

[Ekco]

It actually isn't, I went and scrolled the paper lol

[BradZax]

Originally Posted by Ekco [You must be logged in to view images. Log in or Register.]

It actually isn't, I went and scrolled the paper lol

You're misunderstanding the point of the paper.

I'm not saying these researchers are everyday users.

These are the exact senior scientists at Google DeepMind who build and evaluate Gemini.

The chart isn't supposed to show how a casual user prompts a model; it's a technical evaluation from the actual creators of the AI showing how the underlying architecture behaves.

They absolutely 'matter' because they build the tech.

[Ekco]

My beef is with the METR graph which isn't mentioned once in the paper nor are they, its a sensational Berkeley nonprofit writing disingenuous misleading tests to show the outcome they want, working backwards from ai in scifi scary so we should stop just like they work backward from cows fart too much so you shouldn't be allowed to have a hamburger

[You must be logged in to view images. Log in or Register.]

If you do click one of the dots it does list way past 16hours shown but the methodology of the test itself and the chart are both misleading to show scary exponential growth, nothing changed with how the models themselves are fundementally built, it's stuff around the model that is improving, the harness & MoE. You can stick a earlier model in a harness and let it run for days also but the chart doesn't show that they cap gpt5 at 6 hours and previous models are 12-30 minutes, if you run the same opus they have ranked as they do without a harness it scores will be completely different

This doesn’t mean opus 4.6 can work for ~14 hrs, it means on tasks that would take a human expert ~14 hrs, the agent successfully finishes them 50% of the time. Probably completes them way faster actually

Their either turbo retarded or knowily commiting academic fraud for political/funding reasons, nobody there even works at a frontier lab I assume just nonprofit advocacy fart huffing from what I can tell

https://summify.io/discover/is-ai-ab...-s-not-5GezB1/ one click bait YouTuber to counter another

And the Google fanfic thought expirement about the timeline of one sci fi concept progressing into another sci fi theoretical concept itself I have no issue with

[BradZax]

I posted a document from the lead developers on Gemini, not a youtuber.

Originally Posted by Ekco [You must be logged in to view images. Log in or Register.]

the machine god weighs in

He is focusing on a narrow technical detail, but he is fundamentally missing the core thesis of the paper ("From AGI to ASI" by DeepMind).

Why His Argument Fails

1. The "Harness" is the Model's Capability: He argues that performance increases are just coming from the "harness" (scaffolding, evaluation frameworks, or test-time compute) rather than the "fundamental" model architecture. This is a false dichotomy. Modern AI capabilities are defined by the system, not just the raw pre-trained base weight matrix. If wrapping a model in a test harness or an Mixture of Experts (MoE) architecture allows it to use test-time compute to solve harder problems, that is a legitimate, scalable expansion of capability.

2. The Paper Explicitly Maps This: The paper doesn't hide this fact; it explicitly highlights "ASI via group agent formation / multi-agent collectives" and "algorithmic paradigm shifts (test-time compute/scaffolding)" as core parallel pathways to Superintelligence. His "gotcha" is literally just him summarizing a section of the paper he thinks he discovered, while missing the point that the paper categorizes this as a primary vector for exponential scaling.

3. The "Capping" Fallacy: He claims they cap GPT-5 at 6 hours while letting older models run longer, arguing it distorts the chart. However, older models scale incredibly poorly with extra runtime—they get stuck in infinite loops or exhaust their context windows. Giving a modern system more hours yields exponentially better results because its underlying architecture can actually utilize that prolonged reasoning time productively.

Here, argue with AI about it.

This is the part I liked anyway.

Originally Posted by BradZax [You must be logged in to view images. Log in or Register.]

This will validate all the AI haters so much, so enjoy.

[You must be logged in to view images. Log in or Register.]

[Ekco]

My beef is with the METR graph which isn't mentioned once in the paper

Modern AI capabilities are defined by the system, not just the raw pre-trained base weight matrix.

That's literally what the graph is, knowingly and on purpose comparing raw models vs models with harnesses it's apples and oranges and they graphed it

In early 2026, tech analysts and AI researchers heavily panned METR’s capability timelines. Critics pointed out that METR's data is plagued by basic errors

Maybe ask ai if that's a fair comparison, point 3 is wrong also, you can totally put 5.1 in a harness and it would score higher point 1

He argues that performance increases are just coming from the "harness"

lol wut when did I say the harness is where the performance is coming from, this entire output is garbage I can only imagine what kind of fucked up prompt you put in to get this to be spit out and be this confused, try Claude or Chatgpt not grok or Google overview

His "gotcha" is literally just him summarizing a section of the paper he thinks he discovered

I didn't even read the paper, this is just common ai knowledge in articles and current debates, esp in the AGI/ASI debate people at Google disagree with other people at Google about this.

this is the same company that had a mustard tiger named Blake Lemoine who thought a now obsolete chatbot from years ago had a fucking soul because of his religious views, smart people can be retarded and have views on super ai gonna kill us all or turn the entire universe into paperclips, one dipshit taken seriously until recently was even afraid of the concept of a ASI in the future with time traveling capabilities that would torture him for eternity for not working on ai, these are just thought experiments same as AGI/ASI