Models shift to continuous/iterative training
By this point “finishes training” is a bit of a misnomer; models are frequently updated to newer versions trained on additional data or partially re-trained to patch some weaknesses.
Late 2025: The World’s Most Expensive AI · Archived PDF, p. 4
The two-lab weight-update test is demonstrated by August 21, 2025, before the late-2025 deadline. Confirmed 90% concerns iterative training, not daily online learning or a proven first-ever occurrence.
At a glance
- Assessment: Confirmed
- Confidence in assessment: 90%
- Predicted timing: Late 2025
- Primary source: Late 2025: The World’s Most Expensive AI, archived PDF p. 4
What AI 2027 Predicted
Tracker summary: The scenario describes a shift away from discrete model releases toward continuous, iterative training. Rather than training a model from scratch and releasing it as a finished product, labs would frequently update models, training on additional data, patching weaknesses, and releasing incremental versions. The term “finishes training” becomes a misnomer. Archived PDF, p. 4.
Scenario source locator: Late 2025: The World’s Most Expensive AI: archived PDF, p. 4.
How We Track This
Tracker operational test: At least two frontier labs document weight updates to a deployed model by December 31, 2025, through additional-data training or partial retraining. Record the deployment, dated update, lab disclosure and whether pretraining or post-training weights changed. Product releases, prompt changes and scaffold changes alone do not pass this test.
Confirmed (90%): the two-lab test is demonstrated by August 21, 2025, before the late-2025 deadline. OpenAI documents deployed post-training updates and DeepSeek documents continued pretraining of V3 plus replacement of hosted endpoints. This is a demonstrated-by date, not the first-ever occurrence. Anthropic and Google point-release cadence alone remains insufficient to establish weight lineage.
Current Evidence
August 21, 2025: a second lab documents training and deployment. DeepSeek’s V3.1 announcement describes 840 billion tokens of continued pretraining on V3 for long-context extension. Its dated change log records both deepseek-chat and deepseek-reasoner being upgraded to V3.1. Together with OpenAI’s April GPT-4o post-training update, this satisfies the two-lab weight-update test within 2025. It does not establish daily online learning.
May 2, 2025: primary evidence of iterative post-training: OpenAI’s sycophancy post-mortem describes five major GPT-4o updates since launch, each involving new post-training. It explains supervised fine-tuning followed by reinforcement learning that changes the model’s response probabilities. This directly supports the scenario’s partial-retraining component within the 2025 window; it does not establish continuous pretraining or daily online learning.
February 12, 2026: retrospective interpretation: AI Futures Project’s grading describes GPT-4o, GPT-5 and GPT-5.1 as “probably different continuations of the same base model.” That is the forecasters’ qualified inference, not an OpenAI disclosure of shared base weights. Release frequency alone cannot resolve that lineage.
Dated 2025 point updates: Anthropic released Claude Opus 4.1 on August 5, 2025 as an upgrade to Opus 4, with the dated API snapshot claude-opus-4-1-20250805. Google announced an upgraded Gemini 2.5 Pro preview on June 5, 2025, building on its May version and addressing response-style feedback. These are within-2025 examples of iterative releases, not disclosures that the versions share base weights or use continuous pretraining.
NVIDIA describes continuous post-training as a production feedback loop in which deployment surfaces new problems, repeated runs generate rollouts, rewards are verified, and updated weights flow back into training. NVIDIA also identifies Prime Intellect as continuously post-training frontier open models on Blackwell infrastructure. This provides direct vendor evidence for iterative weight updates after deployment, while leaving continual pretraining and daily frontier-model online learning unproven.
Sources:
- GPT-5 is here: OpenAI
- Introducing GPT-5.2: OpenAI
- Introducing GPT-5.2-Codex: OpenAI
- Grading AI 2027’s 2025 Predictions: AI Futures Project
- NVIDIA: Post-training and intelligence per dollar
Counterevidence & Limitations
- Continued pretraining, post-training weight changes and prompt or scaffold updates are different processes; releases may not identify which occurred.
- The distinction between “continuous training” (as AI 2027 envisions: weight updates from new data) and “continuous deployment” (same base model, improved scaffolding) matters for the deeper claim
- Some version bumps may be more about marketing than fundamental model changes
What Would Change Our Assessment
- Strengthen: Additional labs and longer deployment histories broaden the two-lab evidence already established within 2025; daily online learning remains a separate claim.
- Weaken: Labs disclose that the cited versions were separate training runs with no shared weights, or that supposed training updates changed only prompts or scaffolds.
- Reassess scope: Examples involving only prompts or scaffolds do not support the training component. Post-training can change weights and should not be dismissed as superficial.
Update History
| Date | Update |
|---|---|
| 2026-11-09 | Assessment reviewed. Status: Confirmed; confidence: 90%. OpenAI’s April post-training update and DeepSeek’s August 21 continued-pretraining/deployed-endpoint disclosures establish the two-lab test by August 21, 2025. This predates the late-2025 deadline, without identifying the first-ever occurrence or daily online learning. |
| 2026-11-07 | Correction: Added dated 2025 releases and OpenAI’s May post-training account; the forecasters’ shared-base-model statement remains a qualified inference. The two-lab weight-update test requires documented training changes, not release cadence alone. |
| 2026-09-06 | Current evidence and assessment explanation reviewed. Clarified measurement scope, source interpretation and limitations. |
| 2026-07-20 | NVIDIA described continuous post-training as a production feedback loop with repeated rollouts and updated weights, and identified Prime Intellect as continuously post-training frontier open models. This supports iterative training while leaving continual pretraining and daily online learning unproven. Confidence remains 0.90. |
| 2026-03-13 | Continuous/iterative training now standard practice at frontier labs. AI Futures Project graded this prediction as correct. |
| 2025-12 | GPT-5.2 released — sixth major GPT version in ~8 months. “Code Red” response to Gemini 3 makes explicit that labs iterate continuously in response to competitive signals. GPT-5.x series demonstrates rapid iterative updates, with multiple versions released within months. Claude model cadence follows similar pattern. |
| 2025-11 | GPT-5.1 (Nov 12), GPT-5.1-Codex-Max (Nov 18), Gemini 3 six days later, Claude Opus 4.5 six days after that. Three major updates in under a month. The concept of distinct “model releases” is blurring into continuous iteration. |
| 2025-05 | Claude Opus 4 and Sonnet 4 released together (May 22). Google Gemini 2.5 Flash at I/O (May 20). Three frontier labs shipping major updates within days of each other. |
| 2025-04 | OpenAI releases GPT-4.1, GPT-4.1-mini, and GPT-4.1-nano simultaneously (April 14) while deprecating GPT-4.5 Preview. Multiple model variants at different capability/cost points shipped together signals iterative, continuous development replacing discrete “generation” releases. |