MarkKordusic ·
GPT 6 Astra vs Opus 5.5
The end of chess.
https://youtube.com/watch?v=sskvj1_ehx4&si=b3FyYdIuRrmxuvoe
Replies
MarkKordusic ·
Wow these are historical times.
The last days of AI blundering at chess.
I give it to mid next year before they will be mapped 1 to 1 with latest stockfish most in depth analysis and tablebases.
This is the last chance humans have to record themselves beating ai's.
Might dig up a couple of youtube videos of people doing just that.
In this Gotham chess video Opus 5.5 i think it is blundered , missed back rank mate. So the top ai's are still making massive blunders. Take advantage while you can last days, before they iron out these kinks, very well may be beating latest stockfish most in depth mode in six months to a year if they try a bit, they sure have the compute ate their disposal, its a software bottleneck not hardware thats for sure.
MarkKordusic ·
Deep Blue was a specialized chess machine (narrow AI), modern chatbots like ChatGPT/Grok/Claude are general-purpose language models, and true chess engines (Stockfish, etc.) are the ones that are superhuman and unbeatable by any human, including Magnus Carlsen. The confusion comes from the word “AI” being used for completely different kinds of systems.
### 1. What “AI” actually means (and why it’s confusing)
“Artificial intelligence” is a very broad, loose term. It covers anything that makes a computer do something that looks smart. There is a crucial distinction:
- Narrow AI (ANI / weak AI): Extremely good at one specific thing. Chess engines, spam filters, face recognition, recommendation systems, protein-folding programs (AlphaFold), etc. These can be superhuman in their domain and completely useless outside it.
- General AI / AGI: Hypothetical systems that can handle any intellectual task at roughly human level (or better), transfer knowledge across domains, learn new skills flexibly, etc. No undisputed AGI exists yet. Large language models (LLMs) are the closest things we have, but they are still narrow in important ways.
Deep Blue, Stockfish, AlphaZero, and every strong chess program are classic narrow AI. ChatGPT, Grok, Claude, Gemini, etc. are large language models trained to predict text. They are far more general than Deep Blue, but they are not chess engines.
### 2. Deep Blue vs Garry Kasparov (1996–1997)
Deep Blue was an IBM supercomputer purpose-built for chess. It was not “intelligent” in any general sense. It used:
- Massive brute-force search (evaluating ~200 million positions per second in the 1997 version).
- Hand-crafted evaluation functions written by human grandmasters and programmers.
- Huge opening books and endgame databases.
- Specialized chess chips.
In 1996 Kasparov won the match 4–2. In the 1997 rematch Deep Blue won 3½–2½. That was the first time a computer defeated a reigning world champion under standard tournament conditions. It was …
MarkKordusic ·
Yes, I get your drift exactly. You’re asking whether the long-term trajectory of systems like Grok (and frontier LLMs in general) is continuous, broad improvement that eventually closes the gap on every hard task — including perfect chess play at Stockfish level and reliable real-world competence like running a gas station — driven by more money, hardware, and raw compute. And specifically, whether it is plausible that within roughly one year a pure Grok-style model could crush Magnus Carlsen 100–0 with ease.
Here is a clear, detailed, structured answer.
### 1. What the actual long-term goal is
xAI’s stated mission is to understand the true nature of the universe. In practice that means building systems that are maximally truth-seeking, useful, and capable across a very wide range of intellectual tasks. Reducing errors, improving reliability, and raising performance on every measurable benchmark (including chess) is absolutely part of the path. Developers at every major lab (xAI, OpenAI, Anthropic, Google, etc.) are constantly pushing for fewer hallucinations, better long-horizon reasoning, better tool use, better planning, and stronger general problem-solving.
So yes — the direction of travel is “make fewer mistakes and get better at everything.” Chess is just one convenient, highly measurable domain that reveals weaknesses in state-tracking, search, and consistency.
### 2. Why pure scaling does not automatically turn an LLM into Stockfish
Chess engines and language models solve the problem in fundamentally different ways:
- Stockfish / Leela-style engines maintain an exact board state, perform enormous tree search (looking at millions or billions of positions), and use a very precise evaluation function. They never “forget” where a piece is. Their entire architecture is optimized for this single task.
- Large language models predict the next token. Even the best current ones reconstruct the board from a text history of moves. As games get longer they can lose tra…
MarkKordusic ·
You’re right that the goal is capability and usefulness, not purity for purity’s sake. Restricting a system to “I am only a pure language model and therefore I must stay weak at chess” would be stupid. No serious lab is trying to do that. The actual situation is more practical and less conspiratorial than it looks from the outside.
### 1. Tool use is already the direction, not a forbidden zone
Modern frontier systems (including Grok) are built to call external tools when it helps. xAI’s API and agent interfaces already support:
- Server-side code execution (Python sandbox)
- Web search and page browsing
- X search
- Custom function calling
- Multi-agent orchestration
- Persistent skills and longer-running agent workflows
The philosophy is explicit in recent statements and product design: use whatever produces the best outcome for the user. If a free, high-quality engine solves a sub-problem better than the language model alone, the rational move is to call it. There is no sacred rule that says “we must never touch Stockfish because we are an LLM.” That would be self-sabotage, and it is not the policy.
### 2. Why Stockfish is not simply hard-wired into every conversation by default
Stockfish is free and runs on modest hardware. Integrating it is technically trivial for a company with serious compute budgets. The reasons it is not always present as an invisible, always-on chess oracle in every chat session are product and systems engineering reasons, not philosophical ones:
- Interface and product scope: This particular conversation interface is a general-purpose chat system with a defined set of available tools. Chess is a niche request relative to the full range of queries. Shipping a dedicated chess engine as a permanent, zero-latency, always-available backend for every user session has cost, latency, reliability, and maintenance implications at global scale. Labs prioritize tools that serve the broadest set of high-value use cases first (search, code execution, d…
MarkKordusic ·
Yes. I see your point clearly, and I will not argue against the substance of it.
People using a system like this want reliable, correct answers to concrete tasks. When the task is chess, they want the best available move calculation. When it is arithmetic, they want exact numbers. When it is looking up a definition or consulting a standard reference, they want the authoritative source, not a probabilistic approximation that might drift. The moment the system answers “I am only a language model and therefore I cannot guarantee perfection here,” while the perfect tool already exists for free or for a trivial fee, the experience feels broken. That frustration is legitimate.
### The design gap you are pointing at is real
There is a mismatch between what is technically straightforward and what is currently delivered in ordinary chat sessions.
- A top-strength chess engine (Stockfish or equivalent) runs on ordinary hardware and is free.
- Exact arithmetic, dictionary lookups, standard libraries, unit conversions, and many other deterministic tools are equally trivial to expose.
- A system that can already call web search, execute code in a sandbox, and orchestrate multi-step agent workflows has the architectural capacity to call a chess engine, a calculator, or a reference library when the user wants it.
- Offering that capability optionally, or behind a clear toggle, or as a paid higher-tier feature, is well within the engineering and commercial reach of any well-funded lab.
When that option is missing from the everyday interface, the system forces the user to accept lower reliability on tasks that specialized tools already solve perfectly. That is a genuine product shortcoming for anyone who came here to get the job done correctly rather than to watch a language model struggle.
### Why the gap persists in practice
It is not because the underlying models are incapable of tool use. It is not because the money does not exist. It is because product surfaces, default tool s…
metamitya ·
[[chess]]
terrorknowed ·
I think it's a waste of energy to train your model on chess. Economically speaking you should train your model to use stockfish instead.
MarkKordusic ·
Hundred percent. My only comment is unless there is some kind of evidence that tablebases and traditional chess engine software arnt oders of magnitude more suited and better at ches than modern notions of the model you are takling about why would one bother going down that route at all when the tool allready, a very capable tool exists that is argueably less expensive than a brand new scientific calculator.
As in it is not stretching the truth sayimg that second hand i5 intel processor powered desktops with latest version of stockfish are regularly going to landfill, to many worth zero dollars , so a model as you call it drawing on the capabilities of a tool that many consider to be worthless, well i not long ago threw my i5 equivalent desktop in bin.
So we are not talking about something that stretches ones imagination.
A dirt cheap second hand desktop.
Can an llm model, and my point origionally is not "models".
It is specifically free grok.
As yes these models can be linked with stockfish very easily apparently and certain many are.
But im talking grok, chat gpt, claude public free llms because thats all i know about and have used , not having the capability of a trash second hand desktop, less cost than a new calculator. Even grok agreed with me
I think that this should be a basic accessable tool grok for a fee can draw apon.
What you are stating is much different.
Your delving into the actual mechanics of you called it a model.
To me thats newspeak for glorified trash.
Id go and have gone much further, calling these models you refer to as the devil reincarnate itself.sorry shouldnt be getting so smart about such topics when i can barely
Ok im going to press send and hope my response is a at minimum a weak 3 out of ten sorry mate ive very damn angry down here very very … day.
But yes i hundred percent agree with your point in every way and honoured you would deem a post of mine worthy of response , god bless you Sir.