A Global Workspace in LLMs?
It is quite remarkable that something resembling a global workspace emerged inside Claude during training, rather than being explicitly programmed into its architecture.
Shared-workspace designs themselves are not new. Software engineers have used blackboard architectures, shared memory, message buses and coordinator-orchestrator patterns for decades – and nobody regarded those systems as conscious. What is remarkable here is that gradient descent appears to have rediscovered a related organisational principle.
This may be an architectural analogue of instrumental convergence. There appears to be structure in mind-design space: non-arbitrary “architectural attractors” towards which sufficiently capable learning systems may converge because they provide efficient ways of coordinating complex cognition. Anthropic found that Claude’s J-space is unusually well connected to the rest of the network, supports flexible reuse of information and causally mediates multi-step reasoning. All this without architectural imposition.
None of this establishes sentience or phenomenal consciousness – a subjective “what it is like” to be a Claude model. Global Workspace Theory concerns the global availability of information, often called access consciousness. Most agree that functional access and subjective experience should not be bound together merely because both contain the word consciousness. Anthropic is explicit that its experiments do not show that Claude feels anything.
Even more intriguing is the growing evidence that LLMs develop internal symbolic-like or algorithmic reasoning mechanisms, rather than depending entirely on external symbolic modules. Mechanistic work has identified internal processes resembling symbol abstraction, induction over abstract variables and retrieval. That is a qualitatively richer internal computational organisation than the caricature of progress as merely pattern matching across more and more parameters – although whether it constitutes symbolic reasoning in the full classical sense remains disputed.1
The contrast between Google DeepMind’s 2024 and 2025 International Mathematical Olympiad results is particularly revealing.
In 2025, an advanced version of Gemini with Deep Think solved five of the six IMO problems, scoring 35 out of 42 – that’s gold-medal standard. Google reports that it worked end-to-end in natural language, directly from the official problem statements, within the 4.5-hour competition limit. Unlike the previous system, it did not require human experts to translate each problem into a formal mathematical language before solving it.2 3
In 2024, AlphaProof and AlphaGeometry 2 together solved four of six problems and scored 28 out of 42 – one point below that year’s gold threshold. Human experts first translated the problems into domain-specific formal representations. AlphaProof searched for machine-checkable proofs in Lean using reinforcement learning, while AlphaGeometry 2 combined a neural model with a dedicated symbolic geometry engine. One problem was solved within minutes; others required up to three days.
This comparison does not show that the 2024 systems performed no internal symbolic processing, or that the 2025 Gemini was necessarily more (or less) symbolic internally. It does show a movement from specialist, explicitly neuro-symbolic and formally mediated systems towards a much more general natural-language reasoner.
There is, however, a potentially concerning trade-off.
The 2024 AlphaProof outputs were formal proof objects whose correctness could be mechanically checked by Lean. The 2025 Gemini outputs were informal natural-language proofs assessed by expert human graders. They were genuinely evaluated, not simply accepted on trust, but they lacked the native machine-checkability of formal proofs. DeepMind itself previously noted that natural-language reasoning can produce plausible but invalid intermediate steps, whereas formal languages allow proofs to be verified mechanically.
The concern is not that natural-language mathematical reasoning is undesirable. It is that capability and verifiability can move in opposite directions. The more general and celebrated system was also the one whose submitted reasoning could not be checked automatically by the same formal machinery.
This resembles the verifiable threshold4 running partially in reverse, where capability crossed upwards while the strength of the verification channel moved downwards – formal proof checking to expert human judgement.
That is still manageable at the IMO, where excellent mathematicians can inspect six proofs. It becomes more troubling when AI systems produce thousands of arguments, millions of lines of code, scientific hypotheses or strategic recommendations that exceed the time, expertise or cognitive capacity of their human overseers. Generation may become cheap and superhuman while reliable verification remains slow, expensive and human-bounded.
The danger is that persuasiveness, apparent coherence and benchmark success become substitutes for inspectable correctness. A model may produce an answer that looks more human, more general and more impressive precisely as the route by which we could mechanically establish that it is right disappears.
The desirable endpoint is not a choice between general natural-language intelligence and formal verification. It is a system that can generate understandable informal reasoning while also producing formal certificates, executable checks, reproducible evidence, reasoning trackes or other independently verifiable artefacts.
For a project arguing that verification – not merely eloquence or apparent intelligence – is what ultimately matters, the IMO transition is a useful cautionary example of how auditability can become secondary when capability improves.
Footnotes
- I discussed some of the history of emergent symbolic like reasoning in llms here. ↩︎
- A qualification is necessary: “end-to-end in natural language” does not prove that every relevant computation occurred inside one undifferentiated LLM, nor has Google disclosed enough detail to establish that literally no external infrastructure or tooling was involved. Deep Think itself explores multiple solution paths in parallel. What can reasonably be said is that the published approach did not rely on the 2024 pipeline of human formalisation and specialist formal solvers. ↩︎
- Splitting hairs here? the fact that AlphaProof + AlphaGeometry 2 used external tools and humans doesn’t mean there was no internal symbolic processing going on, or that the newer Gemini models became more, or less, symbol-like. ↩︎
- The verifiable threshold idea is discussed here. ↩︎
