Lede
Smarter is not the same as useful.
Introducing GPT-5
What does not make sense
- Fix the corner, not repaint the house. Models often regenerate everything. Small edit ignored.
- Instruction soup. Add five constraints and watch the format buckle.
- Long thinking sold as wisdom, billed as tokens.
- Energy talk says efficient, the grid says bigger bills.
- New model, same old hallucinations with nicer lighting.
Sense check / The numbers
- Data centres and AI power. IEA projects electricity for data centres rising from about 460 TWh in 2024 to over 1,000 TWh by 2030. AI is the main driver. Date: 10 Apr 2025. [IEA]
- Compute keeps exploding. Training compute for frontier models has grown about 4x to 5x per year since 2010. Top runs now reach around 3.5e26 FLOP. Date: 13 Jan 2025. [Epoch AI]
- Per query power. A typical GPT-4o style query is roughly 0.3 Wh, far below older viral claims, but at scale small watts become big bills. Date: 7 Feb 2025. [Epoch AI]
- GPT-5 claims. OpenAI says GPT-5 is smarter across coding, maths, writing, health, and vision, and knows when to think longer. Date: 7 Aug 2025. [OpenAI]
- Hallucinations are still a class of failure, not a myth. Peer-reviewed work catalogues types and risks across domains. Date: 2024. [Nature]
- Instruction following under load. New 2025 benchmarks show drop-offs as constraint density rises. Dates: 15 Jul 2025 and 3 months prior for related work. [arXiv IFScale, ACL Findings]
- Chain of thought costs. Reports in 2025 show diminishing returns from verbose reasoning, while self-consistency methods improve accuracy at extra compute. Dates: 8 Jun 2025 and mid-2025. [Wharton report, ACL CISC]
The sketch
Scene 1: Data centre corridor. A small tag on a server door: PATCH. A forklift delivers ten identical servers marked FULL REBUILD.
Scene 2: Office wall with sticky note: tiny bug. A robot arm slams a giant button labelled REGENERATE ALL.
Scene 3: User pointing at a single word on a document. Assistant hands over a stack of 10 versions labelled 1 to 10. None match the marked word.
What to watch, not the show
- Incentives. Benchmarks reward showy reasoning, not fast, faithful edits.
- Product defaults. One-shot regen trumps targeted tools.
- Cost curves. Cheaper tokens invite brute force loops.
- Governance by buzzword. Efficiency claims outpace metering.
- Human factors. Multi-constraint asks are common, and models still flinch.
The Hermit take
Count fixes, not tokens.
If it cannot keep a thought for three steps, it is not wise. It is noisy.
Keep or toss
Verdict: Toss.
Keep: small, surgical edit tools and reliable instruction checks.
Toss: the regen fetish for every tiny change.
Sources
IEA energy and AI overview – https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
IEA press note, AI driving data-centre demand, 10 Apr 2025 – https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works
IEA energy supply for AI – https://www.iea.org/reports/energy-and-ai/energy-supply-for-ai
Epoch AI compute trends, 13 Jan 2025 – https://epoch.ai/trends
Epoch AI model scale list, 30 Jan 2025 – https://epoch.ai/data-insights/models-over-1e25-flop
Epoch AI estimate of per-query energy, 7 Feb 2025 – https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
OpenAI, Introducing GPT-5, 7 Aug 2025 – https://openai.com/index/introducing-gpt-5/
OpenAI, GPT-5 product page – https://openai.com/gpt-5/
OpenAI, o1 preview and learning to reason, 12 Sep 2024 – https://openai.com/index/learning-to-reason-with-llms/
OpenAI, o3 and o4-mini, 16 Apr 2025 – https://openai.com/index/introducing-o3-and-o4-mini/
Nature, AI hallucination classification, 2024 – https://www.nature.com/articles/s41599-024-03811-x
ACL Findings 2025, Confidence-Informed Self-Consistency – https://aclanthology.org/2025.findings-acl.1030.pdf
Wharton GAIL report, Diminishing returns from CoT, 8 Jun 2025 – https://gail.wharton.upenn.edu/research-and-insights/tech-report-chain-of-thought/
arXiv 2507.11538, How Many Instructions Can LLMs Follow at Once, 15 Jul 2025 – https://arxiv.org/html/2507.11538v1
ACL Findings 2025, Structured Flow Benchmark for Multi-turn Instruction – https://aclanthology.org/2025.findings-acl.486.pdf
CarbonBrief explainer, 15 Sep 2025 – https://www.carbonbrief.org/ai-five-charts-that-put-data-centre-energy-use-and-emissions-into-context/
One response
This really nails the paradox of progress — every smarter iteration of AI seems to make the actual *work* more convoluted, not less. The bit about ‘instruction soup’ hit home; constraint-heavy prompts often break the very structure they’re meant to clarify. It’s a good reminder that capability doesn’t automatically translate to usefulness without better design discipline on the human side.