“To a first approximation, all the big tech companies, from OpenAI to Google to Meta to xAI to Anthropic to several Chinese companies, keep doing the same experiment over and over: building ever larger LLMs in hopes of reaching AGI.”
Our reading:
To a first approximation, all the big tech companies, from OpenAI, Google, Meta, xAI and Anthropic to several Chinese companies, keep doing the same experiment over and over: building ever larger LLMs in hopes of reaching AGI.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
OpenAI called GPT-4.5 its largest model on February 27, 2025 [e1]. Meta disclosed nearly two trillion total parameters and 288 billion active parameters for Behemoth on April 5, 2025 [e2]. Google linked Gemini 3 to AGI on November 18, 2025, as Qwen did for Qwen3 on April 29, 2025 [e3][e4]. These disclosures establish parts of the claim, but do not establish repeated enlargement toward AGI across the full group [e1][e2][e3][e4]. Google's announcement, xAI's July 9 release, and Anthropic's November 24 release provide no comparable model-size histories [e3][e5][e6]. Qwen, DeepSeek, and Baidu's records do not establish both repeated enlargement and an AGI link for all three companies [e4][e7][e8]. xAI's reinforcement learning and DeepSeek's September 29 architectural change do not establish that either company stopped pursuing larger models [e5][e7].
Caveats
- The evidence does not establish the complete claim for January 1 through November 27, 2025, or a material exception to it.
- Missing size disclosures do not show that models did not grow.
- Release announcements do not provide complete histories of private development programs.
- Total parameters, active parameters, training compute, and output-token efficiency are different measures.
Evidence (8 items)
Supports the claim: OpenAI,
our largest
Why it matters: Directly supports enlargement for OpenAI through GPT‑4.5. It does not establish the entire cross-company claim or independently establish AGI intent.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Supports the claim: Meta,
two trillion total parameters.
Why it matters: Provides direct evidence of Meta building a very large frontier LLM, with total and active parameters distinguished. This supports a component of the claim without establishing every company's repeated enlargement or AGI intent.
- Publisher
- Meta
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Supports the claim: Google,
on the path toward AGI
Why it matters: Explicitly establishes AGI intent for Google's Gemini program shortly before the post. It does not establish that Gemini 3 was larger than its predecessors.
- Publisher
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Supports the claim: Qwen Team, Alibaba,
Artificial General Intelligence
Why it matters: Establishes AGI intent for one Chinese company's LLM program. One such company cannot satisfy the several-companies requirement, and intention alone does not establish repeated enlargement.
- Publisher
- Qwen Team, Alibaba
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: xAI,
reinforcement learning
Why it matters: Shows a substantive training development beyond enlarging a model. It cautions against treating the experiments as literally identical, but does not disprove Marcus's approximate description or establish a size trend.
- Publisher
- xAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Anthropic,
dramatically fewer tokens
Why it matters: Documents an efficiency improvement in Anthropic's latest pre-post model. Output-token efficiency is a different measure from model size and cannot establish either enlargement or shrinkage.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: DeepSeek,
Sparse Attention
Why it matters: Documents an architectural change at another Chinese lab. This qualifies a literal same-experiment reading, but does not show that larger-model development had ceased.
- Publisher
- DeepSeek
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Baidu,
10 distinct variants.
Why it matters: Establishes a third Chinese company's relevant model program. A new model family alone does not establish repeated enlargement toward AGI.
- Publisher
- Baidu
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- to a first approximation (their words)
Where we had to interpret
- The plain reading is that repeatedly enlarging frontier LLMs remained a central, shared route toward AGI across the companies described. Evidence about only some named labs cannot establish the claim.
- The approximation allows differences in architecture, training methods and product offerings. Smaller models and other research programs do not by themselves disprove a continuing strategy of building larger frontier LLMs.
- Larger ordinarily means greater model size, established through comparable parameter counts or explicit company statements about relative size. More infrastructure, training compute, inference effort or benchmark performance alone does not establish larger models.
- Total and active parameters must be distinguished when comparing mixture-of-experts models. Multimodal language models count as LLMs under this ordinary reading.
- For this ongoing observation, the reporting window is the current calendar year through publication: January 1 to November 27, 2025. Earlier generations may supply comparison context. This window is an evaluator choice because the speaker supplied no starting date.
- Several Chinese companies is read as at least three identifiable companies. The speaker did not identify them.
Right if: Public records establish repeated enlargement of frontier LLMs as a central route toward AGI across major AI developers, including every named lab and at least three identifiable Chinese companies. The record must establish model enlargement and AGI intent for the same companies; infrastructure expansion or a few examples is insufficient.
Wrong if: Public records establish a material exception to the claimed shared strategy: a named lab, or a substantial part of the described group, was not pursuing repeated enlargement of frontier LLMs toward AGI. An isolated smaller release, additional research methods, or missing size disclosures is insufficient.
We wrote this test after the outcome was known.
What would settle it
Obtain dated records for January 1 through November 27, 2025, with earlier models as comparison points. Support requires comparable size histories and explicit AGI links showing repeated enlargement as a central strategy at every named lab and at least three identified Chinese companies. Contradiction requires evidence that a named lab or a substantial part of the group did not pursue that strategy. Smaller releases or additional research methods alone would not settle it.
We found 16 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, A trillion dollars is a terrible thing to waste, paragraph 18.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2025-11-27.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, A trillion dollars is a terrible thing to waste
To a first approximation, all the big tech companies, from OpenAI to Google to Meta to xAI to Anthropic to several Chinese companies, keep doing the same experiment over and over: building ever larger LLMs in hopes of reaching AGI.
- Where
- paragraph 18
- Published
- 2025-11-27
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: a cited source didn’t say what the write-up said.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.