OpenAI’s Major Setback: GPT‑5 Is Just a Reskinned GPT‑4o, No Breakthrough After 2.5 Years of Pretraining
After Ilya’s Departure: No Breakthrough in Sight
Report by New Intelligence Source
Editors: KingHZ, Taozi
---
📢 Lead Summary
OpenAI is in urgent need of a comeback.
Today’s biggest leak online: The cornerstone of GPT‑5 is actually GPT‑4o — and since the release of 4o, subsequent pre-training has repeatedly encountered obstacles, to the point of near abandonment.
---
OpenAI Core Pre-Training: Consecutive Setbacks?
Rumors suggest GPT‑5’s foundation is still GPT‑4o, with all pre-training versions after GPT‑4.5 abandoned.
This isn’t idle speculation — the claim comes from a recent SemiAnalysis report revealing:
- Since GPT‑4o’s release, OpenAI’s top-tier team has yet to complete a single, large-scale pre-training run for a true next‑gen frontier model.
---
Hardware & Industry Context
- Google is leveraging TPUv7 to potentially challenge NVIDIA’s CUDA moat.
- OpenAI trains its full-stack models entirely on NVIDIA GPUs.
Yet without large-scale model pre-training, inference and post-training hardware debates mean little — pre-training is the most resource-intensive stage in the entire AI chain.
Insider facts:
- ✔ Google’s TPU successfully met pre‑training challenges.
- ✘ OpenAI has made no progress in pre-training since GPT‑4o’s May 2024 launch.
---
Timeline of Stagnation
- Last 2.5 years: No real scale-up in OpenAI’s pre-training.
- Possible “ceiling” was hit after 4o — coinciding with Ilya’s departure.
Even Ilya commented recently: Scaling won’t stop, but “something important” will still be missing.
---
Major Pre-Training Collapse: The “Orion” Case
- Late last year, “Orion” was intended to be GPT‑5.
- After underperforming, it was downgraded to GPT‑4.5.
Issues with Orion:
- Training took over 3 months — breaking the industry norm of 1–2 months.
- Performance gains were mostly in language, with coding quality worse than older models.
- Costs were higher without proportionate capability boosts.
---
Innovation Pressure & Ecosystem Needs
Frontier model setbacks point to compute limits and ecosystem gaps.
Sustainable AI R&D demands integrated platforms connecting development, deployment, and monetization.
Example: AiToEarn — open-source AI content monetization platform.
It enables creation, publishing, and monetization across:
- Douyin
- Kwai
- Bilibili
- Xiaohongshu
- Threads
- YouTube
- X (Twitter)
With analytics and AI model rankings, AiToEarn bridges AI generation with real-world economic impact.
---
February Origins of GPT‑4.5 (“Orion”)
Focus areas:
- Advanced language capabilities
- Stable conversational experience
- Expanded knowledge base
Keyword for GPT‑4.5: Emotional Intelligence.
Coding ability improved, but less emphasized — supporting rumors that Orion’s leap was modest.
---
GPT‑4o as Scaling Path?
- August release of GPT‑5 framed by Sam Altman as “Ph.D.-level AI towards AGI” — met with mixed reactions.
- Many saw GPT‑5 as a refined GPT‑4.5, not a disruptive overhaul.
Evidence points to no large-scale pretraining of an entirely new model, possibly still based on GPT‑4o.
OpenAI now focuses on reasoning paradigms + RL rather than scaling pretraining — leaving rivals a speed advantage.
---
Altman Acknowledges Google’s Lead
Leaked memos show Altman conceding Google’s excellence in LLM pretraining.
Pretraining Stage Essentials:
- Feed massive data to the model (e.g., web pages).
- Teaches relationships & structure — precursor to post-training and deployment.
---
Why This Matters
- Google’s breakthroughs made Gemini 3 deeper in reasoning.
- Given OpenAI’s 2023 setbacks, this was surprising.
- OpenAI pivoted toward heavy-compute “reasoning models” but these failed to scale successfully.
Javier Alba de Alba on GPT‑5:
> Excellent performance and cost-effectiveness — but far from the generational leap expected.
---
Strategic Shift: Scaling → Reasoning
For creators and developers, this shift reinforces the need for adaptability.
Platforms like AiToEarn官网 connect AI outputs to monetization and analytics, allowing reach across multiple channels.
---
GPT‑5: Improvements but Not a Leap
- Boost in coding
- Advanced reasoning
- Hallucination reduction
- Enhanced medical applications
- Unified naming (no more turbo/mini/o-series labels)
Javier Alba de Alba warns: GPT‑5 is closer to “GPT‑4.2” than a new-gen product.
Epoch AI suspects GPT‑5 used less training compute than GPT‑4.5.
---
Future Plans: “Shallotpeat”
Goal: Fix persistent pre-training issues.
For users: GPT‑5 is good but evolutionary, not revolutionary.
---
Platform Strategy Insights (Sherwin Wu Interview)
Shift in consensus:
- From “one supreme model” → Specialized diversified models (e.g., Codex, Sora).
- Separate tech stacks for text, image, video.
Fine-Tuning Evolution:
- Early: adjust tone/instructions
- Now: Reinforcement Fine-Tuning (RFT) using proprietary enterprise data.
---
Agent Development Model
Dual-track strategy: App + API to reach both consumers and developers.
Two agent types:
- Undirected / Exploratory
- Procedural / SOP-oriented — where hard-coded logic is essential.
Agent Builder provides controlled, verifiable agent workflows — critical for regulated industries.
---
One More Thing: December Launch Wave
Rumored releases:
- Image Gen v2
- IMO & IOI multimodal models
- GPT‑5.2 Codex
---
Reference
Leaked memo: Sam Altman sees ‘rough vibes’ and economic headwinds at OpenAI
---
Instant ASI Tracking
⭐ Like · Share · Watch — all in one click ⭐
Turn on notifications to get instant updates from New Zhiyuan.
---
Shallotpeat could be OpenAI’s consolidation before a major leap — aligning talent, compute, and API innovation to match Gemini’s pace.
Meanwhile, creator economy tools like AiToEarn let AI-generated content go global in one click. For developers tracking these shifts, integration of model performance, ecosystems, and monetization will be key in the months ahead.
---
Would you like me to now produce a concise timeline infographic mapping OpenAI’s pre-training milestones from GPT‑4o to “Shallotpeat”? That would make these developments even clearer.
