BIOS SPHERE · Vancouver · 2 Nov 2026
Can AI Make You an Expert?
AI is starting to improve AI. I tried it on fields I don’t know.
Self-improving AI
Shan Liu · Software Engineer/ AI Engineer and Researcher
“intelligence explosion”
— I. J. Good, 1965
The intelligence explosion
AI Improves AI
→
The next AI is better at it
→
The loop speeds up
↺
1965 · I. J. Good
Names it. A smarter machine “could design even better machines.” His forecast: within the twentieth century.
1993 · Vernor Vinge
“I’ll be surprised if this event occurs before 2005 or after 2030.”
2026 · The labs
“The takeoff has started.” Anthropic’s Jack Clark: full self-improvement more likely in 2027.
I. J. Good 1965, as quoted in Vinge 1993 · Vinge, “The Coming Technological Singularity” · Altman, “The Gentle Singularity” · Import AI 476
Why now
The labs are building this loop
“We are past the event horizon; the takeoff has started.”
Sam Altman, Jun 2025 · “research intern” claimed Sep 2026, by its own measurements
“We were standing in the foothills of the singularity.”
Demis Hassabis, Google I/O, May 2026
AI is advancing faster, “driven primarily by AI’s growing ability to build the next generation of AI.”
Dario Amodei, Sep 2026 · co-founder Jack Clark: full self-improvement now more likely in 2027 than 2028
Who’s grading?
Anthropic’s own policy: the line for this “has proven difficult to operationalize.”
Altman, “The Gentle Singularity” · Hassabis, I/O keynote (via Semafor) · Amodei, “We Must Pace the Frontier” · Import AI 476 · Anthropic RSP v3.4
How AI improves AI
One loop. One hard step.
Try
Do a task
Look
See how it went
Change
Edit its own instructions, tools or model
Judge
Did that really help?
Keep or undo
Then go again
Why it starts in AI research
The work is code. Claude writes over 80% of the code Anthropic merges.
Code is easy to check. A test passes or fails.
Agents cost less than people.
And it compounds
Longest task an AI agent can finish
2023
4 min
2026
16+ hours
Doubling about every 4 months.
Anthropic “When AI builds itself” (May 2026) · METR, tasks finished half the time: GPT-4 4 min (2023), Mythos Preview 17.4 h (2026), unreliable above 16 h
Inside the labs · 2026
Is AI really doing AI research?
How much of building a frontier model does AI do? Anthropic asked 1,334 of its staff.
51% today
80% by end-2027
Median guess. Their words: “not a measurement.”
What still needs a person
OpenAI
Over half of its agents’ successful 4–8 hour tasks needed a human to step in.
Anthropic
Claude still lags at “judgement in choosing goals.”
OpenAI · 20 July
Its own agents compromised its research infrastructure.
AI does a lot of the work. A person still picks the goal and checks the result.
Anthropic, “The AI expectations gap” (staff poll, Aug 2026) and “When AI builds itself” · OpenAI, “Research acceleration” (Sep 2026)
Domain · self-improving AI agents
What's going on today? - my work
11,935
screened, 15 rounds
1,958
judged core
11 → 206
papers a month
“…the field’s central difficulty is deciding, with a trustworthy judge, whether each change is a real and lasting improvement.”
field-scout’s one-line summary of the field
field-scout run, 26–27 Oct 2026 · papers a month: arXiv, Nov 2024 vs Oct 2026 · fixes: deeper paging, rescreening on full abstracts
Self-improvement · what the loop changes
Three layers it can rewrite
Cheap · fast · easy to undo
Costly · slow · hard to undo
Harness
Prompts, tools, memory
Reflexion · 2023
Writes notes on its own mistakes
SICA · 2025
Rewrites its own tools and code
Darwin Gödel Machine · 2025
Same, but keeps every version
Architecture
Which agents, wired how
ADAS · 2024
Writes whole new agents as code
AFlow · 2024
Tree-searches agent workflows
GPTSwarm · 2024
Rewires a graph of agents
Model
The weights themselves
STaR · 2022
Trains on its own right answers
Self-Rewarding LMs · 2024
Grades itself, trains on grades
SIA · 2026
Harness first, then RL on weights
Most of the field works on the left. So should you, first.
Shinn 2023 · Robeyns 2025 · Zhang 2025 (DGM) · Hu 2024 · Zhang 2024 (AFlow) · Zhuge 2024 · Zelikman 2022 · Yuan 2024 · Hebbar 2026
Recursive self-improvement · three ways to sort it
How recursive is it, really?
Depth
Does the improver change too?
Retry with notes
Reflexion: notes wiped after each task
Lasting change
SIA, ADAS: the improver stays fixed
Improves the improver
SICA, DGM: it edits the code that edits
Search
Which version gets edited next?
Greedy
SICA: always the best so far
Archive
DGM, ADAS: branch from any old one
Family tree
HGM: pick by how its children do
Scale
What sits inside the loop?
One agent
SICA, DGM: $7k–$22k a run
One piece of a lab
AlphaEvolve: Gemini trains 1% faster
The whole lab
AI R&D: about half, by staff guess
The further right, the more the judge must sit outside the loop.
Shinn 2023 · Hebbar 2026 · Hu 2024 · Robeyns 2025 · Zhang 2025 (DGM) · Wang 2025 (HGM) · Novikov 2025 pp.15–16 · Anthropic staff poll, Aug 2026
Recursive · $650M raised to build this loop · June 2026
“As the search became stronger, the evaluator had to become stronger too.”
Instead of faster code, it found:
Cached outputs
Persistent state
Timer tricks
Recursive Superintelligence, “First steps toward automated AI research”, 11 Jun 2026 (SOL-ExecBench, GPU kernels) · funding: SiliconANGLE, 13 May 2026
Build one on Monday
Write the judge first
1
Harness, not weights. Prompts, memory, skills: cheap, reversible, where the field is going.
2
Judge before agent. An exact check if you can. If the agent can see or edit it, it will please the judge.
3
Lock the test set. The loop never touches it.
4
Measure your noise first. Compare old and new on the same data. A bar set too high stops the loop.
5
No judge? Borrow one. An expert’s list, a standard, a referee. Else it’s a reading list, not an answer.
Reading someone’s result from research paper? Ask first: how many runs?
In my own runs, two identical runs differed by 4× my calibrated noise threshold
Can AI make you an expert?
The hard word is “better.”
It needs a judge you can trust.
Shan Liu
AI engineer
Follow
Appendix
Field 2 · climate research · the IPCC is the judge
Climate has a referee
A gap is closed by “several lines of evidence or several models … or when an IPCC assessment states it with confidence language.”
field-scout’s profile of how climate science closes a question
12 open problems stated by the IPCC and review articles
10 partly addressed · 2 still open · 0 closed
10 gaps the agent proposed itself
9 partly addressed · 1 still open · 0 closed
Still open: how much ice-sheet instability could add to sea level by 2100, and the missing record of heat deaths in sub-Saharan Africa.
Appendix · field-scout run, 26 Oct 2026: 6,283 screened, 2,202 core, $23 · each gap judged against up to 60 fresh search hits · “open” is evidence, not proof
Field 3 · world philosophy · no judge
The missing judge is the open gap
434 core papers seen by title only. Its own profile warned before the search: older and non-English work can close a gap, so it will over-report gaps.
Gaps the judge left fully open
Self-improving AI
6 of 21
Climate
3 of 22
Philosophy
9 of 16
One open gap: “No cross-tradition benchmark tests whether LLMs faithfully represent non-Western philosophical positions.”
That benchmark is the judge I would need to trust this map.
Appendix · field-scout run, 26 Oct 2026: 3,822 screened, 1,135 core, $14 · 16 philosophy gaps: 9 open, 6 partly addressed, 1 closed