1---
2slug: 001-start-here-field-overviews
3renderings: []
4audience: public-docs
5name: 'Research Blog #1: Start Here - Four Maps of AI Safety, and How Researchers See It'
6question: What do the four whole-field overviews of technical AI safety (the Shallow Review, Open Phil's research directions, the AI Safety Field Map, and the Foundational Challenges agenda) each cover, where is each skewed, where do they disagree about which areas matter, and which areas do this site's reviews already cover? And how do individual researchers explain the problem, and which question does each view hinge on?
7status: published
8date: '2026-10-11'
9summary:
10 thesis: 'Read the four field overviews as answers to four different questions: the Shallow Review maps who works on which agenda, Open Phil''s list says what one funder wanted to pay for in early 2025, the Field Map counts people and money by organization, and the Foundational Challenges agenda lists the problems still open for LLMs as of 2024. They disagree most on interpretability, theory, evals and adversarial robustness, and that disagreement is partly an artifact of how each one counts.'
11refs:
12 interest:
13 - url: https://w.sh/interest/practical-alignment
14---
15
16## Read Four Maps First, Each Answering a Different Question
17
18Every review on this site starts from four overviews of the whole field. They are not four versions of one map: each answers a different question, and none is current.
19
20The Foundational Challenges agenda was posted on 15 April 2024, the oldest of the four. The Shallow Review covers work up to 30 November 2025 and Open Phil's list dates from its early-2025 request for proposals. The Field Map's data run is dated 30 June 2026, but most of its headcounts date from 27 September 2025, about a year ago.
21
22| Overview | Made by | What it covers | Current as of |
23| -- | -- | -- | -- |
24| [¶ Shallow Review of Technical AI Safety, 2025](https://w.sh/paper/leech-2025-shallow-review-technical-ai) | • Arb Research, nine authors led by [Gavin Leech](https://w.sh/person/gavin-leech)<br>• Third yearly edition | • About 80 research agendas plus one card per frontier lab<br>• About 800 papers and posts<br>• Each card lists names, estimated FTEs, funders and critiques | Work from 30 November 2024 to 30 November 2025 |
25| [¶ Research Directions Open Phil Wants to Fund in Technical AI Safety](https://w.sh/paper/tais-rfp-research-areas) | • Open Philanthropy, now Coefficient Giving<br>• [Jake Mendel](https://w.sh/person/jake-mendel), Max Nadeau and Peter Favaloro | • 21 research areas in 5 clusters<br>• Example projects and key papers per area<br>• Stars mark favourite areas; daggers mark a high bar | A USD 40M request for proposals, open until 15 April 2025 |
26| [¶ AI Safety Field Map](https://w.sh/paper/waterman-2026-ai-safety-field-map) | Harry Waterman | • 172 organizations in 10 subfields (plus political action committees)<br>• Headcount, budget and 508 tracked grants | • Data run dated 30 June 2026<br>• Headcounts mostly as of 27 September 2025 |
27| [¶ Foundational Challenges in Assuring Alignment and Safety of Large Language Models](https://w.sh/paper/anwar-2024-foundational-challenges-assuring-alignment) | • 42 authors led by [Usman Anwar](https://w.sh/person/usman-anwar)<br>• Mostly academic | • 18 challenges in 3 parts: scientific understanding of LLMs, development and deployment methods, and sociotechnical challenges<br>• More than 200 concrete research questions | Posted 15 April 2024 |
28
29- **The Shallow Review describes supply**: which agendas exist and who works on them.
30 - It deliberately does not say which ones work.
31 - Its authors call it shallow because "we are not specialists in almost any of it" and each entry got about two hours.
32- **Open Phil's list describes demand**: what one funder wanted to pay for.
33- **The Field Map describes scale**: how many people and dollars sit in each organization.
34- **Foundational Challenges describes open problems**: what is still unsolved, phrased as research questions.
35
36None of the four covers anything after mid-2026, and the three research maps stop in 2025 or earlier. Events the organization reviews here lean on, such as the 2026 changes at the frontier labs, are newer than every map.
37
38## Each Map Has a Skew to Correct for Before Trusting It
39
40### The Shallow Review Gives Six Iterative-Alignment Cards One Inherited List of Assumptions
41
42The Shallow Review's "Iterative alignment" group (alignment at pretraining and post-training time, black-box make-AI-solve-it, inoculation prompting, in-context learning, and inference-time steering) carries one parent theory of change, which each of the six cards on the website displays as its own. It opens "LLMs don't seem very dangerous and might scale to AGI, things are generally smooth, relevant capabilities are harder than alignment", then strings together assumptions such as "assume no mesaoptimisers" and "assume that alignment is a superficial feature".
43
44That field is a list of assumptions the authors attribute to the whole group, not a theory of change any of the six methods states for itself. Read it as the authors' summary of what a skeptic would say iterative alignment presumes. The same review estimates more than 500 FTEs on iterative alignment "in some sense", the largest FTE estimate in its data file.
45
46The numbers come from the review's own data file of 16 December 2025 in its [code and data repository](https://github.com/arb-consulting/shallow-review-2025), which carries 74 agendas and 5 lab cards.
47
48### Open Phil's List Is One Funder's Wish List, and That Funder Pays for Most of the Field
49
50Open Phil's guide calls itself "an opinionated overview" and ranks its own areas:
51
52- **Starred, its "favorite topics" (8)**:
53 - Four in adversarial machine learning: jailbreaks and unintentional misalignment, control evaluations, backdoors and alignment stress tests, and alternatives to adversarial training.
54 - Two in sophisticated misbehavior: alignment faking, and encoded reasoning in chains of thought.
55 - Two white-box: applications of white-box techniques, and white-box estimation of rare misbehavior.
56- **Daggered, "a high bar" (3)**: more transparent architectures, conceptual clarity about risks, and new moonshots for aligning superintelligence.
57- **Sent elsewhere**: capability evaluations and governance go to separate requests for proposals, so their absence here is scope, not a verdict.
58
59The funder behind the list dominates the money the Shallow Review and the Field Map measure. Of the roughly USD 874 million in grants with a stated amount on the Field Map, Coefficient Giving (Open Phil's new name) gave about USD 733 million, 84 percent, and Good Ventures, its main backer, another 4 percent (summed from the grant rows in the map's page data).
60
61The Shallow Review names Open Phil or Coefficient Giving as a funder on 26 of its 67 "Funded by" lines, counted from its 16 December 2025 [data file](https://github.com/arb-consulting/shallow-review-2025). Coefficient Giving publishes its grants, so a tracker built from public grant pages over-counts it; that is this review's inference, not the map's.
62
63### The Field Map Files Each Lab's Whole Safety Team Under One Subfield
64
65The Field Map sizes subfields by summing organizations, and each organization sits in exactly one subfield. Anthropic's 40 safety FTEs and Google DeepMind's 30 are all filed under interpretability, and OpenAI's 15 under LLM safety. AI control therefore shows only 16 FTEs (Redwood Research's 15 and one at Luthien), while the Shallow Review estimates 5 to 50 FTEs on control across organizations.
66
67Budgets carry a similar distortion. Alignment theory shows USD 66.9 million, but USD 39.9 million of it is one entry, ARIA, listed with 1 FTE. The map itself says to treat its figures "as directional, not authoritative".
68
69The two FTE sources are not independent either: Stephen McAleese co-authored the [¶ Shallow Review](https://w.sh/paper/leech-2025-shallow-review-t