“Instead, OpenAI will be throwing the kitchen sink at future efforts to build GPT-5 , including “test-time-compute”, a new approach that involves longer, more expensive inference times rather than “constant time inference” of GPT-4-like models, as well as (I suspect) massive amounts of synthetic data”
Our reading, the sentence that was checked:
OpenAI will throw the kitchen sink at future efforts to build GPT-5, including test-time compute and, Marcus suspects, massive amounts of synthetic data, which appears to work better for semi-closed domains like math and coding than in the open-ended real world.
Assessment: Supported Settled before its deadline on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Breaking: OpenAI's efforts at pure scaling have hit a wall., paragraph 15.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-breaking-openais-efforts-at-pure-a7f675-a2
We recorded 10 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-02-13. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- Settled before its deadline
- Deadline state on 2026-10-29
- Settled before its deadline (the release’s reference date)
Conditions
- the synthetic-data part is a suspicion ('I suspect') (qualifier, stated by the speaker)
- the domain comparison is hedged ('appears to work better') (qualifier, stated by the speaker)
- future efforts to build GPT-5 (scope, stated by the speaker)
From the speaker
- throwing the kitchen sink at future efforts to build GPT-5
- including “test-time-compute”
- longer, more expensive inference times
- (I suspect) massive amounts of synthetic data
- appears to work better for semi-closed domains like math and coding
Our interpretation
- The firm forecast is that GPT-5 will incorporate test-time reasoning and multiple approaches beyond pure pretraining scaling.
- Read 'throwing the kitchen sink' as an ordinary expression for combining approaches, without requiring every possible technique or a specified number of techniques.
- Read test-time compute as allocating additional computation to reasoning before answering. The wording supplies no numerical latency or cost threshold and does not require every query to use extended reasoning.
- The synthetic-data suspicion and hedged domain comparison are excluded from scoring entirely.
- GPT-5's public release and accompanying technical documentation can establish whether the forecast development approach was implemented.
Counts as supported if: GPT-5 release documentation establishes that the released system incorporates additional inference-time reasoning and other methods beyond pure pretraining scaling.
Counts as contradicted if: Definitive documentation establishes that GPT-5 was developed and released using pure pretraining scaling without the forecast additional inference-time reasoning. Silence about methods alone would not establish contradiction.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- Settled before its deadline
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
GPT-5's August 7, 2025 release resolves the firm forecast. OpenAI documented adjustable thinking time and reinforcement-learning training for its reasoning models. These establish test-time reasoning and a combination of approaches beyond pure pretraining scaling. Selective use of reasoning is consistent with the claim. The synthetic-data suspicion and hedged domain comparison are not scored. This finding does not establish the post's broader assertion that pure scaling had failed.
Limitations
- The evidence consists of OpenAI's own release and technical disclosures.
- The assessment establishes the use of additional reasoning computation, not a measured increase in inference costs relative to GPT-4.
- The idiom supplies no objective test of how exhaustive OpenAI's development efforts were.
Evidence (3 items)
Supports the claim: OpenAI,
control GPT‑5’s thinking time
Why it matters here: The release documentation describes controllable thinking time in the deployed GPT-5 API. This directly supports the forecast inclusion of additional inference-time reasoning.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: OpenAI,
trained to reason through reinforcement learning.
Why it matters here: The system card documents reinforcement learning for the released reasoning models. Together with e1, this establishes a combination of methods beyond pure pretraining scaling.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: OpenAI,
think longer
Why it matters here: The dated public release announcement describes extended thinking as an implemented GPT-5 capability. It also qualifies the scope: longer reasoning is selective, rather than necessary for every response.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Breaking: OpenAI's efforts at pure scaling have hit a wall.
Instead, OpenAI will be throwing the kitchen sink at future efforts to build GPT-5 , including “test-time-compute”, a new approach that involves longer, more expensive inference times rather than “constant time inference” of GPT-4-like models, as well as (I suspect) massive amounts of synthetic data
- Where
- paragraph 15
- Published
- 2025-02-13
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-breaking-openais-efforts-at-pure-a7f675-a2: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-breaking-openais-efforts-at-pure-a7f675-a2.
HeldBy this site
The openai review found the criteria unfair to the speaker's words.
About assessment marcus-breaking-openais-efforts-at-pure-a7f675-a1, which was not published.