Next Three Months AI Forecast | Generated by AI
Question: Predict the next 3 months (from July 24, 2026)
Answer:
For the next 3 months (late July → late October 2026), my prediction based on current AI trends, your trajectory, and the market:
1. AI models: smaller + cheaper frontier models accelerate
High probability
The biggest shift will not be “bigger models only”, but:
- better inference efficiency
- more MoE
- better post-training
- agent architectures
- specialized models
Expect:
- 30B–100B open models approaching today’s frontier quality for many tasks
- more local AI deployment
- more companies optimizing cost/token
The DeepSeek lesson continues:
Algorithmic efficiency becomes a bigger moat than raw GPU count.
Likely winners:
- DeepSeek
- Qwen
- Kimi/Moonshot
- OpenAI
- Anthropic
2. GPU market: VRAM becomes the bottleneck
Your recent GPU discussions are exactly aligned with where things go.
The next 3 months:
- 16GB consumer GPUs become increasingly constrained
- 24GB VRAM becomes the practical minimum for serious local AI
- used RTX 3090 market remains interesting
- AMD ROCm improves but CUDA ecosystem remains dominant
The important metric changes:
2023:
TFLOPS
2026:
usable VRAM + memory bandwidth + software stack
A $1000 GPU with 48GB VRAM may beat a $2000 GPU with 24GB for AI developers.
3. AI coding agents explode
This is probably the biggest near-term change.
Expect:
- Claude Code style agents everywhere
- OpenAI Codex-style workflows
- terminal-native agents
- autonomous repo modification
The programming workflow:
Before:
human
↓
write code
↓
compile/test
After:
human
↓
specification
↓
agent
↓
code + tests + debugging
↓
human review
Your direction with:
- ww
- iclaw
- CLI agents
- agent workflows
is very aligned.
4. Open source AI infrastructure grows
Expect more importance around:
- inference engines
- compilers
- kernels
- distributed serving
Examples:
- vLLM
- SGLang
- TensorRT-LLM
- ROCm
- Triton kernels
The “CUDA monopoly” discussion will continue, but replacement happens gradually.
The stack:
Model
↓
Framework
↓
Compiler
↓
Kernel
↓
Hardware
becomes the battlefield.
5. AI startups: many small teams become possible
I expect more “2-10 person AI companies”.
Because:
2020 startup:
10 engineers
backend
frontend
infra
ML
2026 startup:
2 founders
AI agents
cloud GPUs
open models
The bottleneck moves from coding to:
- finding valuable problems
- distribution
- proprietary data
6. Your personal trajectory prediction
Given your recent experiments:
- GPT-2 training
- ROCm exploration
- GPU kernels
- agents
- AI consulting direction
I expect your next 3 months likely look like:
Month 1
More infrastructure learning:
- PyTorch internals
- CUDA/ROCm
- inference optimization
- vLLM/SGLang
Possible:
- another GPU purchase decision
Month 2
A more concrete AI product:
Examples:
- vertical agent
- AI workflow SaaS
- enterprise automation
- AI consulting demo
Month 3
More public proof:
- GitHub projects
- blog posts
- technical writeups
- customer acquisition
Your strongest advantage is not competing with OpenAI training a 1T model.
It is:
Frontier AI knowledge
+
engineering ability
+
shipping ability
+
cheap compute
That combination is rare.
My probability ranking for the next 3 months
| Event | Probability |
|---|---|
| AI coding agents become mainstream | 90% |
| Open models improve significantly | 85% |
| VRAM shortage continues | 80% |
| More AI startups emerge | 80% |
| CUDA replaced quickly | 20% |
| A new DeepSeek-level efficiency breakthrough | 40% |
| Consumer local AI adoption grows | 70% |
The biggest opportunity window is probably AI agents + cheap inference + domain workflows, not training another foundation model from scratch.
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋