How’s China Doing?

• overall: china’s growth is slowing
• property:
  • despite government support, property prices still struggling
  • falling demand => tighter financing => unfinished projects
• prices are going down—deflation!?

Involution (Neijuan)

• Dropping household wealth => more cautious purchasing.
• Young unemployment rate still elevated

Major Challenges

1. demogratic change
2. changing the growth model: state sector restructuring seems to have stopped
3. shifting to more domestic demand driven by consumers
4. changes in financial system—how can banks avoid over-lending
5. changes in fiscal system (how to find sources of revenue)
6. reducing corporate debt, LGFV debt, etc.

options are not great

• older model has had diminshing returns
• exports and trade surplus may not last
• more limited access to technologies and foreign direct investment
• high debt levels requires many years of debt logistics
• changes to population structure / aging limits consumptino and savings (so increasing domestic demand is hard)
• any fast shift to new growth model will bring a bit of depression

New Direction

1. maintain: focus on investment driven by state industrial policy
2. shift: investment from housing sector to optimized manufacturing (high tech)
3. expand: manufacturing capacity overall—but efficiency
4. double-down: on exports even if tariffs

ambitious and a huge gamble

Three layers of agent safety

1. model architecture: fundamental limitations of transformer structure
2. architecture -> LLMs: training data (poisoning), training objective (reward hacking)
3. LLMs -> prompts: prompt injections, unintended actions, goal scheming

prompt injections

OWASP top 10 for LLM applications…. RAG/Agents are WORSE because humans do not have choice. Web agents, can browse the web and have context poisoning.

evaluation setup

1. etiologic validity
2. realistic threat models
3. systematic evaluations (e.g., obviously anecdotal works)
4. controlled environments

computer security principles

• confidentiality (don’t infiltrate passwords)
• integrity (don’t nuke important files)
• availability (don’t bring things down)

benign inputs leading to harms

• triggering compaction => failures

Unintentional behavior: “unsafe agent behavior that deviations from user intentions from a task”

• SU-SOC175 JAN052025
• SU-SOC175 JAN122025
• SU-SOC175 JAN142025
• SU-SOC175 JAN212025
• SU-SOC175 JAN262025
• SU-SOC175 JAN282026
• SU-SOC175 FEB022025
• SU-SOC175 FEB042025
• SU-SOC175 FEB112026
• SU-SOC175 FEB232026
• SU-SOC175 FEB252026
• SU-SOC175 MAR022026

Equality constrained smooth optimization problem:

\begin{align} \min_{x}\quad & f\qty(x) \\ \textrm{s.t.} \quad & Ax = b \end{align}

for \(f\) convex, and twice differentiable; for \(A \in \mathbb{R}^{p\times n}\), rank \(p\).

additional information

equality constrained quadratic minimization

say its a quadratic:

\begin{align} f\qty(x) = \frac{1}{2} x^{T}P x + q^{T} x + r \end{align}

for \(P \in \mathbb{S}^{n}_{+}\)

We can form optimality via the KKT Conditions in a block:

\begin{align} \mqty(P & A^{T}\\ A & 0) \mqty(x^{*}\\v^{*}) = \mqty(-q \\ b) \end{align}

constraint

recall constraint; our general constraints means that we can select \(f\) within a feasible set \(x \in \mathcal{X}\).

active constraint

an “active constraint” is a constraint which, upon application, changes the solution to be different than the non-constrainted solution. This is always true at the equality constraint, and not necessarily with inequality constraints.

types of constraints

We can write all types of optimization problems into two types of constraints; we will use these conventions EXACTLY: