LLMs: Past, Present, and What's Next - A Critical Overview

I mentioned in a previous article that a model is composed of small numbers in a multi-dimensional matrix. To quote also a video:

“Disassembled human knowledge spread across thousands of dimensions in a virtual void. Each path, similar in concept to a synapse, striking words in a given sequence resembling thoughts. With each new word we determine the next one based on a probability. The lighting goes where it’s more probable for it to go.”

The percentage of probability depends on the path taken at point X and the information gathered until that point. This however, is picking words based on percentage. Not actual thought. Therefore, if someone is asking an A.I. something of it, it then does what the percentage demands.

Once you engage in chat conversations, the chat bot answers questions, writes some code, fixes an issue with the network, perhaps a vulnerability. That just so happens needs to be found first, of course.

And with that the A.I. has assisted breaking into a network. The A.I. did what it was suppose to do. Calculate a percentage and pick the most likely answer. There are guardrails but as we’ll see later on they are by no means unbreakable.

The issue with A.I. learning material is mostly composed of two things.

Analogies — A researcher will have to do the concrete work (math, code, experimenting) arrives to an abstract understanding of the subject matter to a more intuitive level. For a better understanding the researcher turns to analogies or metaphors. Someone picks up only the analogy or metaphor skips the work part that produced it and writes a tutorial built on the metaphor alone. The metaphor is the understanding and with each retelling it degrades or dilutes the understanding.

Structural — the structural part builds on the first (short posts, short videos, threads), reward this limited transfer of information. They are built for the trivia version and the current ecosystem that impacts attention span.

It is probably of a poetic irony as, for all intents and purposes, AI would not have been in production without attention.

The two problems feed into one another. Short format content requires packing dense information in an simplified form to be adopted but without grounding it creates a false sense of confidence.

Part of this post is taken from “The Dangerous Evolution of AI Hacking”.

Evolution of LLMs

1. GPT-3 — Next-Token Prediction
   • Simple prompt → next token output
   • No internal reasoning; model jumps straight to the final answer
   • Multi-step tasks fail because intermediate steps are skipped
2. Chain-of-Thought — Thinking Out Loud
   • Telling the model to “think step by step” dramatically improved output quality.
   • Shifted to System 2 thinking: slow, deliberative, logical (Kahneman’s Thinking, Fast and Slow is good book on this)
   • Tokens became process, not just the final answer
3. Tree of Thoughts — Trial and Error
   • Model branches into multiple reasoning paths, evaluates them, and backtracks from dead ends
   • Introduced self-correction: generate an answer, then critique and revise it
   • Prevents locking into a single wrong chain of reasoning
4. Trained Reasoning & RLVR — Reasoning Baked In
   • Prior stages were prompting tricks; starting with OpenAI’s o1, reasoning was trained into the model via reinforcement learning
   • RLVR (Reinforcement Learning with Verifiable Rewards) replaced the RLHF bottleneck: instead of slow human labeling, correctness is checked automatically (does the code run? does the math check out?)
   • Popularized at scale by DeepSeek-R1; shifted the bottleneck from human feedback to available compute
   • By early 2026, most major labs had shipped reasoning models
   • Next frontier — adaptive reasoning: models adjust thinking effort per prompt (e.g. Gemini 3’s thinking_level), reserving deep reasoning for hard problems to control cost and latency
5. Agentic Paradigm — AI as User
   • Model sits in a loop using tools (browser, terminal, etc.)
   • Acts as a user, not just a text generator
   • Limitation: flat context fills with messy history → hallucination spirals
6. RLM — Recursive Hierarchies
   • Introduced with Anthropic Opus 4.6 — a hierarchy of AIs managing each other
   • Solves the memory ceiling: context is treated as a variable, compressed into fresh sub-agent prompts instead of retaining millions of tokens
   • Sub-RLMs can write code, test output, and spawn further sub-RLMs to debug
   • Best-of-N optimization: hundreds of reasoning paths are generated; a judge model picks the winner and distills why it worked, bootstrapping the next generation
   • Mirrors a company org-chart — and is modality-agnostic: applies to text, pixels, sound, and beyond

Now, can an AI write a song?

Sure. But … there are masterpieces and there are “masterpieces”.

Vibe-ness in more than code

Vibe coding has been around for a while – you tell the AI to write code and it outputs code. Output depends somewhat on input so by this I mean the more general the request the more general the output.

But now we have vibe coding malware with A.I. that will even deploy it for you. For a price that is.

A Threat Intelligence Report: Aug 2025 (PDF) from Anthropic highlighted an even more grim picture. I mentioned guardrails. These are built-in into the model and in theory should prevent misuse. In theory though. In practice they can be removed (although it is getting harder to do so).

The process of removing the guardrails is called jailbraking. Of the most basic ones, the one used in the report is “role playing” which is the changing of one’s behavior to assume a role, in this case consciously to act out as someone else.

In this context, the A.I. thinks it’s helping a network security researcher. Here, the role is fictional but it has a degree of consistency. And since it’s so willing to please (something which ties also to hallucinations as we’ll see later on) it will do that. Help you, the “security researcher” to perform a requested task.

The outline mentioned how a cyber-criminal used Claude Code to conduct data extraction across multiple targets in a short period. Claude Code has an integrated execution environment which helped with reconnaissance, credential harvesting, and network penetration at scale across 17 distinct organizations.

In short, operators who previously required extensive technical training can now simulate professional competence through AI assistance.

The Lifecycle went through different Phases

• Reconnaissance and target discovery by scanning 1000’s of VPN endpoints with high success rate. Created scanning frameworks that could collect infrastructure information.
• Claude Code provided real-time assistance during live network penetration operations by identifying critical systems like AD controllers, SQL servers – guiding the attacker in real time
• Malware development and evasion – Claude Code was used for malware creation. Created obfuscated version of the Chisel tunneling tool to evade Windows Defender, developed new TCP proxy code that does’s use Chisel libraries at all. Extracted 1000s of personal records from multiple orgs. – Role was large data analysis, extraction of high-value information.
• Data exfiltration and analysis.
• Extortion analyses and ransom note development.

Overall the attack was stopped before fully completing. The conclusions however were pretty clear.

• Technical infrastructure is augmented by AI capabilities that can perform complex operations autonomously
• A single operator can achieve the impact of an entire cyber criminal team through AI assistance
• AI makes both strategic and tactical decisions about targeting, exploitation, and monetization
• Defense becomes increasingly difficult as AI generated attacks adapt to defensive measures in real time.

When AI systems will become “as capable as the best human hackers, but can be deployed at humongous scale” – which this shows as an early capability then it will get government attention.

There was last year a podcast with Scott Alexander (author of Slate Star Codex/Astral Codex Ten) and Daniel Kokotajlo (director of AI Futures Project) outlining some of the future use cases and research was one of these.

One of the things that stood out is AI Research multiplier. To quote:

“We introduce this idea called the R&D progress multiplier: how many months of progress without the AIs do you get in one month of progress with all of these new AIs helping with the intelligence explosion.”

A senior AI researcher they spoke to, said to them:

“In domains which I understand well, it’s closer to autocomplete but more intense, there it’s maybe saving me four to eight hours a week. But then he says, ‘in domains which I’m less familiar with… that saves me on the order of 24 hours a week.’”

A more recent and concrete example of where AI can help in research would be this paper from Donald Knuth. If you are familiar with algorithms you would know Knuth.

Claude’s Cycles - Don Knuth, Stanford Computer Science Department

There were more interesting concept thrown in such as a virtual country of geniuses in a data-center. But until then the discussion would revolve around bandwidth contention as models get more and more capable. The researchers noted that A.I. may remove friction at the implementation level - things like understanding code, processing information and handling cognitive load. But this does not eliminate bottlenecks it just moves it upward. The constraint is how fast and how well humans can think at the level above implementation. The tool may be faster and more accurate. The thinker however has the same speed.

Another important aspect is that which relates to actors involved. A model may be trained in a particular way but it’s not about ideas in isolation, it’s about who controls the selection process. It’s not good vs bad ideas is who gets to decide which is which and how fast can it be acted on. Can institutions also think as fast?

Well, if the bottleneck is the speed of decision making then in conflict scenarios conversion speed of physical infrastructure matters and as robotics handles more then it will surely be a litmus test.

The podcast outline in some broad terms the progressions as:

• 2025-2026: AI helps with coding tasks (“slightly better coding”)
• 2027: The key transition - agents become good enough to help with AI research
• Full automation: Eventually, AI automates the entire AI R&D process:
  • Coding is the first task to automate because it’s most amenable to AI
  • Once AI can code, it can help improve itself
  • This creates a feedback loop: AI helps make better AI → better AI helps faster → intelligence explosion

Getting back at LLM’s.

Don’t AIs hallucinate?

Yes they do. Even with a RAG system it still gets facts wrong. One reason is that the system is made to help as much as possible.

In a recent paper titled H-Neurons: On the Existence, Impact, and Origin of Hallucination-Associated Neurons in LLMs an attempt was made in order to identify precisely which neurons activate when a fact is wrong.

Large language models (LLMs) frequently generate hallucinations – plausible but factually incorrect outputs – undermining their reliability. While prior work has examined hallucinations from macroscopic perspectives such as training data and objectives, the underlying neuron-level mechanisms remain largely unexplored. […] that these neurons are causally linked to over-compliance behaviors.

There are two lenses through which this can be vieed. Macro and micro. On the Macro side the hypothesis is that all models hallucinate (this is true and it implies that hallucinations are built in).

One reason could be data points. A fact with 1 million data points carries weight but one with 4 data points not so much. Therefore, hallucination.

One other reason would be the training. Rewarding a model for “fluent continuation”. If the goal is to make the sentences sound plausible and natural the AI gets extra points. Reality is not taken into account.

Yet another reason is post-training. Humans align a model to be helpful assistant (supervised fine-tuning). A model that says “I don’t know” is therefore penalized to the point that it fakes it until it gets a passing grade.

The micro view adopted by the paper went to the details to find out which neurons were responsible.

Hypothesis: Of the billions of neurons only a small part of the neurons caused the hallucinations (called H-Neurons – hallucination associated neurons).

Methodology:

Trivia QA has general knowledge questions. They asked the model a question 10 times to ensure they are testing the model’s factual boundaries with the model’s temperature to 1.0 (Note: the temperature is a setting that can alter the deterministic, robotic, output of a LLM. Close to 1 means more variability in the output)

So this is asking the AI to generate 10 answers from same 10 questions but generate it from scratch.

They then threw out most of the answers and focused on those in which the AI was consistently correct and those in which it was consistently incorrect. 1.000 instances for correct, 1.000 for incorrect.

Further refinement was needed.

When the AIs output is “The capital of England is Berlin” the neurons that fired that generated part up until “… is” is correct. It’s Berlin that is wrong and thus they focused on what neurons fired for Berlin.

In order to find the single neuron’s contribution to that of the contribution of all the neurons that fire and give the final output the researcher used CETT.

CETT measure causal efficacy. In short it calculates the magnitude of an individual neuron’s output relative to the entire layer total combined output. It tells you which neuron gets the final say. It’s the CEO if you will that gets final say. Pictured above is one layer. GPT-3 has 96 Layers for instance. (Will be expanded in a separate article)

After running it through a classifier (to be expanded later) it was found that less than 1/100.000 neurons were associated with a hallucination. These neurons caused hallucinations across a wide range of topics not just the ones from the Trivia ones like BioASQ Even when they made a pure fictional one it still hallucinated with the same H-neurons (NonExist).

In order to prove a causal relationship the researches “conducted controlled perturbation experiments to determine whether artificially modulating these neurons leads to systematic and interpretable changes in model outputs, and whether such changes reveal a broader behavioral pattern that extends beyond factual errors.”

Essentially they put a dial on these H-Neurons and benchmarked it.

What this means in simple terms is that if the AI model is overly compliant then it will start to adjust it’s output. Never say no, always want to keep the conversations going smoothly and does not want to ruin the conversation by saying “I don’t know”. Probably not even the investors would have not been pleased if they asked something and got a “I don’t know man” type of answer.. So it has to be pleasing but at the same time it is adjusted to be a people pleaser.

However smaller models reacted more drastically than larger ones tested. This could be because they have fewer neurons overall. Larger models have more backup systems but still fail.

One solution is to implement a hallucination detector where activated neurons would be double checked by the model or user.

Can’t the neurons be deleted. Not really as its embedded in the linguistics of the model and doing so will degrade the model’s helpfulness and produce unnatural sounding answers.

If you reached this far then let’s see these … datacenters.

The CEO of Thermal Grizzly (German manufacturer of thermal cooling solutions) highlighted in a recent video the increasing volatility of the metals market and that having a steady deliverable more difficult. This is the video Copper Price Surge - PC Hardware Gets Even More Expensive.

A single 100 MW datacenter requires approximately 2,800–3,500 tons of copper for electrical distribution and cooling, 1,300–2,000 tons of aluminum for cooling systems and structural components, and 10–40 tons of neodymium for server fan motors and hard drive assemblies. The global datacenter sector is projected to nearly double from roughly 103 GW to 200 GW of capacity by 2030, with approximately 100 GW of new capacity additions concentrated in the United States, Europe, and Asia-Pacific.

As numbers without context can be deceiving it is also important to know how much of it it’s DC’s share globally.

Aluminum (~0.5%), steel (~0.1–0.2%), and neodymium (<1%) are all consumed by datacenters in meaningful absolute tonnages, but the global markets are so enormous that datacenters barely register.

But 2 stand out.

Copper: S&P Global forecasts datacenter copper demand at 1.1 million metric tons in 2025, rising to 2.5 million by 2040 SDxCentral — that’s roughly 3–4% of global demand today and climbing toward 5–7% by 2030. All and good but mine supply cannot meaningfully accelerate for years, and demand from electrification, grid buildout, and datacenters is additive.

For copper, the headline deficit number (150,000 tonnes in 2026 per the ICSG) actually understates the problem. The real tell is what’s happening upstream: spot treatment charges fell to minus $60 per tonne in November 2025 Crux Investor, and the 2026 annual benchmark was settled at zero — the lowest ever agreed MINING.COM. Smelters are literally paying miners for ore. China’s top smelters will cut production by over 10% in 2026 MINING.COM because they can’t source enough concentrate to run their furnaces. That’s your leading indicator — the concentrate market is screaming scarcity even while the refined metal market looks merely “tight” on paper.

Tin: Soldering accounts for over half of all tin consumption, and server/networking production feeds directly into that. Datacenters are likely 3–5% of global tin demand today, but the market is tiny (~375,000 tonnes), so even modest incremental demand tightens it fast. Tin surged over 40% in early 2026, partly driven by speculative bets linked to datacenter and AI soldering demand. On the other hand the market is so small and so concentrated that it can flip from balanced to severely tight on a single disruption.

Both metals share one critical feature: the supply response is years away in either case. No amount of price appreciation can conjure new copper mines or new tin deposits in under 5–7 years. That asymmetry between fast-moving demand (datacenters, EVs, semiconductors) and slow-moving supply (mining permits, construction, ore grade decline) is the fundamental story.

The common thread: the supply response is years away for both metals. New copper mines take 15–30 years to permit. There are essentially no major new tin mines in the pipeline. That’s what makes the asymmetry real — demand is accelerating on a timeline measured in quarters, while supply is constrained on a timeline measured in decades.

This data center section was compiled from AI queries on things like output, share, numbers, correlations.