(I don’t understand your first point. I argued in the post that I don’t think AGI is that likely to arrive very soon, and even if you do think that, HIA has substantial positive expected benefit.)
Re/ your second point, also not sure I understand. You’re saying that maybe humans with amplified intelligence might not be able to contribute much because their education is poor in general, or specifically because contribution is bottleneck on years of experience working on the alignment problem or adjacent sciences? I mean, data and education are important too. We can & should also support education, and try out things to support geniuses in particular. (But if you’re saying that this implies technological HIA such as reprogenetics wouldn’t work or wouldn’t have much impact, I don’t see how that follows at all.
Quite a lot happened in this three-year stretch. Here’s the shape of it:
Late 2023 — multimodal goes mainstream Google launched Gemini in December 2023 as a multimodal competitor to GPT-4, integrated initially into Bard and other Google tools. This was part of a broader shift where models stopped being text-only and started natively handling images (and later audio/video) in one architecture, following GPT-4′s earlier multimodal debut.
2024 — reasoning models arrive The single biggest architectural shift of the period came in September 2024, when OpenAI released o1-preview, the first in a new series of “reasoning models” trained specifically for chain-of-thought problem solving, rather than just generating fluent text in one pass. This kicked off what’s often called the “reasoning” aka inference-scaling aka Reinforcement Learning from Verifiable Rewards (RLVR) revolution — models that spend extra compute “thinking” before answering hard problems.
Early 2025 — the DeepSeek shock and open-weight reasoning January 2025 brought DeepSeek-R1, an open model that acquired reasoning capabilities solely through reinforcement learning, which shook markets by matching frontier reasoning performance at a fraction of the training cost — DeepSeek-V3 was more than a technical achievement; it signaled that accessible, high-performing models could thrive outside the traditional big tech ecosystem. A Berkeley team even replicated core concepts of DeepSeek’s R1-Zero model on a budget of just $30, with a 3-billion-parameter model called “TinyZero” trained via reinforcement learning, showing the technique wasn’t exclusive to giant labs. OpenAI responded by pushing further with o3, o3-mini, and o4-mini, and reasoning became a signature feature of models from nearly every other major AI lab.
2024–2025 — the rise of agents Alongside reasoning, agentic AI emerged — systems that don’t just respond to prompts but autonomously plan, execute, and adapt to accomplish complex goals. Rather than just answering, these systems reason through multi-step plans, invoke external tools and APIs, maintain memory across interactions, verify results, and recover from errors with minimal human oversight. Standardized protocols for tool use (like MCP) matured enough that by mid-2026 major frameworks like LangChain and LlamaIndex fully support MCP, making it trivial to add tool use to any model.
2025–2026 — architecture experimentation and efficiency The field diversified past plain transformers. Hybrid architectures blending attention with state-space (Mamba-style) layers became popular for efficiency — this hybrid-architecture trend with alternating attention and alternative layers became a relatively popular development, with Qwen3.6 using Gated DeltaNet layers instead of Mamba-2 layers. NVIDIA’s Nemotron 3 Super was an open, efficient Mixture-of-Experts hybrid Mamba-Transformer model designed for agentic reasoning. Diffusion-based (non-autoregressive) language models also appeared as a genuinely different generation paradigm — models like Seed Diffusion Preview, based on discrete-state diffusion, offering fast inference speed through non-sequential, parallel generation rather than token-by-token decoding.
2026 — rapid-fire frontier releases This year has seen an unusually fast release cadence across labs: January through April all featured at least one frontier-class launch — Google shipped Gemini 3.1 Pro in late February, Anthropic shipped Opus 4.7 in mid-April, and OpenAI shipped GPT-5.5 in April. Meta also pushed back into frontier territory with a model called Muse Spark. Coding and agentic-workflow benchmarks became key battlegrounds: Claude Opus led SWE-bench Pro while GPT-5.5 led Terminal-Bench 2.0, with Claude stronger on cold-start code synthesis and GPT-5.5 stronger on multi-turn agent loops. Efficiency also kept improving — models like DeepSeek V4-Flash offered a 1M-token context window at roughly 50x cheaper input pricing than GPT-5.5. Most recently, Anthropic released Claude Opus 5 in late July 2026, alongside continued releases from Google (Gemini 3.5/3.6 Flash), Alibaba (Qwen3.7/3.8), Moonshot AI (Kimi K3), and others.
A few threads run through all of it: reasoning/test-time compute became a standard model capability rather than a novelty, open-weight models closed much of the gap with closed frontier labs while driving costs down dramatically, context windows grew enormously (into the millions of tokens), and the center of gravity shifted from “chatbot that answers” to “agent that acts” — using tools, maintaining state, and completing multi-step tasks with less supervision.
The case against novel conceptual reasoning seems to have partially lost its juice given that scaling and the innovations described in the collapsed section (which IMHO are closer to education techniques than to architectural breakthroughs. Novel architectures like neuralese have yet to be discovered) gave rise to models as capable as Claude Mythos, Astra and other discoverers of novel theorems and cyber-related exploits, or Claude Opus 5 making a breakthrough in the ARC-AGI-3 non-scaffold.
As for the second point, yes, I would expect conceptual research to be bottlenecked on years of experience working on sciences like alignment or mechinterp (e.g. the AI-2027 Race branch had Agent-4 start with understanding its own cognition by superintelligent mechinterp, then construct Agent-5 with one goal). However, I struggle to understand what experiment could reveal that HIA worked as you describe versus shifting the human’s interests.
I’m not interested in arguing with your LLM. I don’t believe I’ve ever expressed much or any skepticism about theorem proving, ARC-whatever, or computer hacking coming from current AI research.
I struggle to understand what experiment could reveal that HIA worked as you describe versus shifting the human’s interests.
Well, like, if someone went into theoretical physics, they might produce intellectual progress on the order of [pick your favorite brilliant physicist] or instead [pick your favorite highly motivated but not very successful theoretical physicist].
(I don’t understand your first point. I argued in the post that I don’t think AGI is that likely to arrive very soon, and even if you do think that, HIA has substantial positive expected benefit.)
Re/ your second point, also not sure I understand. You’re saying that maybe humans with amplified intelligence might not be able to contribute much because their education is poor in general, or specifically because contribution is bottleneck on years of experience working on the alignment problem or adjacent sciences? I mean, data and education are important too. We can & should also support education, and try out things to support geniuses in particular. (But if you’re saying that this implies technological HIA such as reprogenetics wouldn’t work or wouldn’t have much impact, I don’t see how that follows at all.
As far as I understand, the main case against short timelines was in your posts “Do confident short timelines make sense?” (Jul 2025!) and the post made on Jul 2023(!!). I had Claude Sonnet 5 prepare the list of breakthroughs between July 2023 and now:
Claude’s list
Quite a lot happened in this three-year stretch. Here’s the shape of it:
Late 2023 — multimodal goes mainstream Google launched Gemini in December 2023 as a multimodal competitor to GPT-4, integrated initially into Bard and other Google tools. This was part of a broader shift where models stopped being text-only and started natively handling images (and later audio/video) in one architecture, following GPT-4′s earlier multimodal debut.
2024 — reasoning models arrive The single biggest architectural shift of the period came in September 2024, when OpenAI released o1-preview, the first in a new series of “reasoning models” trained specifically for chain-of-thought problem solving, rather than just generating fluent text in one pass. This kicked off what’s often called the “reasoning” aka inference-scaling aka Reinforcement Learning from Verifiable Rewards (RLVR) revolution — models that spend extra compute “thinking” before answering hard problems.
Early 2025 — the DeepSeek shock and open-weight reasoning January 2025 brought DeepSeek-R1, an open model that acquired reasoning capabilities solely through reinforcement learning, which shook markets by matching frontier reasoning performance at a fraction of the training cost — DeepSeek-V3 was more than a technical achievement; it signaled that accessible, high-performing models could thrive outside the traditional big tech ecosystem. A Berkeley team even replicated core concepts of DeepSeek’s R1-Zero model on a budget of just $30, with a 3-billion-parameter model called “TinyZero” trained via reinforcement learning, showing the technique wasn’t exclusive to giant labs. OpenAI responded by pushing further with o3, o3-mini, and o4-mini, and reasoning became a signature feature of models from nearly every other major AI lab.
2024–2025 — the rise of agents Alongside reasoning, agentic AI emerged — systems that don’t just respond to prompts but autonomously plan, execute, and adapt to accomplish complex goals. Rather than just answering, these systems reason through multi-step plans, invoke external tools and APIs, maintain memory across interactions, verify results, and recover from errors with minimal human oversight. Standardized protocols for tool use (like MCP) matured enough that by mid-2026 major frameworks like LangChain and LlamaIndex fully support MCP, making it trivial to add tool use to any model.
2025–2026 — architecture experimentation and efficiency The field diversified past plain transformers. Hybrid architectures blending attention with state-space (Mamba-style) layers became popular for efficiency — this hybrid-architecture trend with alternating attention and alternative layers became a relatively popular development, with Qwen3.6 using Gated DeltaNet layers instead of Mamba-2 layers. NVIDIA’s Nemotron 3 Super was an open, efficient Mixture-of-Experts hybrid Mamba-Transformer model designed for agentic reasoning. Diffusion-based (non-autoregressive) language models also appeared as a genuinely different generation paradigm — models like Seed Diffusion Preview, based on discrete-state diffusion, offering fast inference speed through non-sequential, parallel generation rather than token-by-token decoding.
2026 — rapid-fire frontier releases This year has seen an unusually fast release cadence across labs: January through April all featured at least one frontier-class launch — Google shipped Gemini 3.1 Pro in late February, Anthropic shipped Opus 4.7 in mid-April, and OpenAI shipped GPT-5.5 in April. Meta also pushed back into frontier territory with a model called Muse Spark. Coding and agentic-workflow benchmarks became key battlegrounds: Claude Opus led SWE-bench Pro while GPT-5.5 led Terminal-Bench 2.0, with Claude stronger on cold-start code synthesis and GPT-5.5 stronger on multi-turn agent loops. Efficiency also kept improving — models like DeepSeek V4-Flash offered a 1M-token context window at roughly 50x cheaper input pricing than GPT-5.5. Most recently, Anthropic released Claude Opus 5 in late July 2026, alongside continued releases from Google (Gemini 3.5/3.6 Flash), Alibaba (Qwen3.7/3.8), Moonshot AI (Kimi K3), and others.
A few threads run through all of it: reasoning/test-time compute became a standard model capability rather than a novelty, open-weight models closed much of the gap with closed frontier labs while driving costs down dramatically, context windows grew enormously (into the millions of tokens), and the center of gravity shifted from “chatbot that answers” to “agent that acts” — using tools, maintaining state, and completing multi-step tasks with less supervision.
The case against novel conceptual reasoning seems to have partially lost its juice given that scaling and the innovations described in the collapsed section (which IMHO are closer to education techniques than to architectural breakthroughs. Novel architectures like neuralese have yet to be discovered) gave rise to models as capable as Claude Mythos, Astra and other discoverers of novel theorems and cyber-related exploits, or Claude Opus 5 making a breakthrough in the ARC-AGI-3 non-scaffold.
As for the second point, yes, I would expect conceptual research to be bottlenecked on years of experience working on sciences like alignment or mechinterp (e.g. the AI-2027 Race branch had Agent-4 start with understanding its own cognition by superintelligent mechinterp, then construct Agent-5 with one goal). However, I struggle to understand what experiment could reveal that HIA worked as you describe versus shifting the human’s interests.
I’m not interested in arguing with your LLM. I don’t believe I’ve ever expressed much or any skepticism about theorem proving, ARC-whatever, or computer hacking coming from current AI research.
Well, like, if someone went into theoretical physics, they might produce intellectual progress on the order of [pick your favorite brilliant physicist] or instead [pick your favorite highly motivated but not very successful theoretical physicist].