Sign in

angela zhou

@angelamczhou.bsky.social
7.9K followers 1.1K following 647 posts

assistant prof at USC Data Sciences and Operations and Computer Science; phd Cornell ORIE. data-driven decision-making, operations research/management, causal inference, algorithmic fairness/equity bureaucratic justice warrior angelamzhou.github.io

PostsRepliesMedia
angela zhou @angelamczhou.bsky.social · 5h
Offbeat ask. Are there any teaching materials for teaching students to recognize good ideas? I'm thinking sort of the value of meetings for pitching various ideas until you land something that hits with the room. if execution is free, recognizing what makes an idea good is more important
470
Reposted by angela zhou
Katie Bergh @katiebergh.bsky.social · 10h
75% of the Louisiana families who lost SNAP after the enactment of the Republican reconciliation law were cut off for procedural reasons. Families are losing food assistance due to paperwork issues, not because they no longer meet eligibility requirements. veritenews.org/2026/10/07/f...
Many states, including Louisiana, have placed pressure on applicants to submit more information more often to verify their eligibility to reduce the state’s payment error rate and limit how much the cost of benefits shifts to the state.

“It’s been a race to drive down those error rates,” Llobrera said.

He said the challenge of implementing new program restrictions and requiring more documentation has increased the burden on applicants and state agencies. The agencies have to process more paperwork on a tight timeline without enough staff. This can increase the likelihood of errors when processing applications, potentially leading to rejection of benefits for more people. It’s also harder for applicants to meet additional verification requirements.
1229
angela zhou @angelamczhou.bsky.social · 05/10/2026
john wilson is incredible because it's just data and me reading 160k of casenotes to ground-truth stuff for reviewer comments is also just data and we're all just mainlining data out here
030
Reposted by angela zhou
Kristina Gligoric @IC2S2 @gligoric.bsky.social · 02/10/2026
Check out our new paper! arxiv.org/abs/2603.21404
arxiv.org
Multi-Perspective LLM Annotations for Valid Analyses in Subjective Tasks
Large language models are increasingly used to annotate texts, but their outputs reflect some human perspectives better than others. Existing methods for correcting LLM annotation error assume a singl...
2173
Reposted by angela zhou
Keith Kurson @keithlaugh.love · 03/10/2026
nyc had a separate system for 3rd party social workers to use, that did not connect to the system that the cities social workers used, and it was someones job to migrate data between the two
1141
Reposted by angela zhou
Liam Dillon @liamjdillon.com · 02/10/2026
Years ago I interviewed Houston Mayor Annise Parker who told me one of the main reasons the city was able to cut homelessness by so much is she forced all the governments and nonprofits into a room and made them use the same data system. Meanwhile in LA www.latimes.com/california/s...
The agencies use different data systems, and even different Google spreadsheets, to produce their reports and often have different bureaus within them managing different housing programs.For example, a senior director for Bass told the interviewers that the Inside Safe program uses an internal Google spreadsheet to manage its occupancy records and does not have access to the Homeless Management Information System, the countywide database that tracks street contacts with homeless people and housing placements
9549119
Reposted by angela zhou
Maria Antoniak @mariaa.bsky.social · 02/10/2026
#TADA2026 is now up on lea.ac! Full schedule, filterable by people you follow, a custom feed, and more! (This is not part of the official TADA organization, just an experiment. Try it out and let us know what you think! Also a nice way to follow along virtually if you can't attend in person!)
A screenshot from Lea showing the conference page for TADA. The program tab is selected, and presentations are filtered by "People I follow." Each resulting presentation includes the title, presenters, location, time, a topic tag, and the person (or people) presenting whom I follow.
1217
Reposted by angela zhou
Suresh Venkatasubramanian @geomblog.bsky.social · 03/10/2026
I know it's a trend to quit big tech and write a warning note. But this one from David Robinson is worth reading if only because he is much more careful in where he locates the concerns (about culture and industry attitude to safety) www.theatlantic.com/technology/2...
theatlantic.com
I Quit OpenAI Because Its Culture Is Broken
The industry’s approach to safety will guarantee more failures unless something changes.
193
Reposted by angela zhou
Maria Antoniak @mariaa.bsky.social · 03/10/2026
Maybe worth saying aloud that I do worry about safety risks (hacking, bio, an “out of control” agent). I just also see those risks as deeply entangled with money, politics, personalities, regulation, such that I doubt we can “solve safety” if we don’t first address foundational issues.
3688
Reposted by angela zhou
Hunter Owens @hunterowens.net · 02/10/2026
Incredible set of designs for Sunset Blvd coming from @ladotofficial.bsky.social - cannot wait, thanks @cd13.lacity.gov Eunisses Hernandez and the whole team on an awesome proposal
4163
angela zhou @angelamczhou.bsky.social · 02/10/2026
according to the Other Place, David Robinson (head of policy planning and safety systems lead) left openai the day the safety folks got fired ... I don't know much, but it is worrisome. Only met him a few times at Cornell, but got the sense he cared about humans first and foremost
190
Reposted by angela zhou
Katie Bergh @katiebergh.bsky.social · 01/10/2026
SNAP update: Another cut from the Republican reconciliation law takes effect today, cutting the federal match for SNAP administrative costs from 50% to 25%. Every state now gets half as many federal dollars to ensure SNAP gets to eligible families on time & in the right amount.
11821
angela zhou @angelamczhou.bsky.social · 29/09/2026
I love Jev. Jev is an example of what will be the next "frontier": compiling intelligence to cheaper and more structured outputs. Thinking of running my new AI evals class on Jev or open-source alternatives bc my institution keeps asking us to do AI stuff without additional resources to do so
171
Reposted by angela zhou
Melanie Mitchell @melaniemitchell.bsky.social · 28/09/2026
👀 Florida attorney general to OpenAI: (www.myfloridalegal.com/sites/defaul...)
1140872
angela zhou @angelamczhou.bsky.social · 27/09/2026
280
Reposted by angela zhou
Neil Shephard @neilshephard.bsky.social · 25/09/2026
Harvard Statistics now has a bluesky account: @harvardstats.bsky.social good to follow. Neil.
021
Reposted by angela zhou
professor tools for commensality (red scare edition) @inquiline.myatproto.social · 24/09/2026
one thing bsky doesn't have is trade press
042
Reposted by angela zhou
Carlos Scheidegger @cscheid.net · 23/09/2026
Louder for the people in the back: Jev can’t be calibrated www.alexmolas.com/2026/09/23/j... Friends don’t let friends lie about the accessibility of the DGD even in principle
alexmolas.com
Jev can't be calibrated
Jev is a useful zero-shot classifier, but its probabilities can't be calibrated for your data. Calibration depends on your data distribution, which Jev never sees, so treat its outputs as scores and r...
44410
Reposted by angela zhou
David Mimno @dmimno.bsky.social · 22/09/2026
And I'm going to bet that a lot of what AI companies imagined as the potential exponential-growth big-model use cases aren't writing code or emails, but automating business processes. And Jev/Laya/whatever-comes-next are going to make that vastly simpler and cheaper.
1153
Reposted by angela zhou
David Mimno @dmimno.bsky.social · 22/09/2026
It's possible for Jev/Laya/Decision Models to be not that big a deal as tech and massive as a new paradigm. Here's why I'm really excited from an NLP history perspective (thread)
17519
Reposted by angela zhou
Eugene Vinitsky 🍒 @eugenevinitsky.bsky.social · 20/09/2026
This seems like a reasonable and straightforward norm? and important to impose quickly given that literal stalwarts of the field are seemingly choosing to defect
041
Reposted by angela zhou
Johan Ugander @jugander.bsky.social · 20/09/2026
My Yale colleague Dan Spielman has started a substack, which I thought I'd signal boost since Dan is an outstanding thinker and communicator (and person!) but isn't on social media much. Here's his first substantive post from a month ago: mathadjacent.com/p/llms-testi...
mathadjacent.com
LLMs, Testing, Grade inflation, and Accommodations
Like every professor, I’ve been worrying about how AI systems will force us to change our teaching practices.
1228
Reposted by angela zhou
Sasha Costanza-Chock @schock.cc · 18/09/2026
I had also kind of forgotten that we have this cute video explainer of our paper! Check it out ;) www.youtube.com/watch?v=bH5T...
youtube.com
Who Audits the Auditors?: Recommendations from a field scan of the algorithmic auditing ecosystem
YouTube video by Algorithmic Justice League
076
Reposted by angela zhou
Alondra Nelson @alondra.bsky.social · 19/09/2026
In a new report, "Artificially Inevitable," the NYC Public Advocate calls for the state to enact transparency and consent policies for AI use with personal data, increase corporate liability and deliver an #AIBillofRights advocate.nyc.gov/press/artifi...
advocate.nyc.gov
ARTIFICIALLY INEVITABLE: NYC PUBLIC ADVOCATE RELEASES NEW REPORT ON THE ROLE OF AI AND NECESSARY GUARDRAILS IN NYC
Press Release, September 14th, 2026
04115
Reposted by angela zhou
rain 🌦️ @sunshowers.io · 19/09/2026
Style: Towards Clarity and Grace has so much invaluable advice, such as putting old ideas before new ones. I think that's the sort of thing a lot of experienced technical writers pick up on over time anyway, but it transformed the way I write
1194
angela zhou @angelamczhou.bsky.social · 16/09/2026
Another blog post: separatedhyperplane.substack.com/p/we-cant-le... We can't let frontier labs choose their own auditors - or their own questions. Triangulating between the near-past lens of algorithmic accountability, the financial crisis, and the AI evaluation landscape as it looks now:
separatedhyperplane.substack.com
We can’t let frontier labs choose their own auditors - or their own questions.
We should start with mandatory audits that adequately cover harms under current laws and regulations while we figure out the rest.
2249
Reposted by angela zhou
Katie Bergh @katiebergh.bsky.social · 14/09/2026
Between October 2024 & March 2026, Arizona denied more than half of the households applying or recertifying for SNAP for failing to complete the required interview. But it wasn't for lack of trying: during that period, more than 3 million calls to the state's overwhelmed call center were dropped.
02514
Reposted by angela zhou
Kevin Reuning @reuning.bsky.social · 14/09/2026
I'll post more about this tomorrow, but I've spent the last few weeks trying to do some social network analysis on the most recent "OpenAI Swarm" You can do some interesting things with the data to identify how the different "Agents" interacted. kevinreuning.com/blog/ai_swar...
kevinreuning.com
OpenAI Swarm – Kevin Reuning
095
Reposted by angela zhou
spacecowboy @spacecowboy17.bsky.social · 13/09/2026
Last week I ran an A/B test in For You where the treatment arm had a higher penalty for popular items and items with more paths leading to them were boosted more. Statistically significant results: 0.6% more users liked at least something on a given day 4.7% more likes per user
A table of experiment results:
Metric	Control	Treatment	Difference	% Change
Users	58,367	57,747	-620	-1.1%
Requests	5,596,792	5,657,823	+61,031	+1.1%
Requests per user	95.89	97.98	+2.09	+2.2%
Total likes	3,304,534	3,427,841	+123,307	+3.7%
Liked % (per request)	21.54%	21.65%	+0.11%	+0.5%
Likes per request	0.5904	0.6059	+0.0155	+2.6%
Total show more	13,386	12,199	-1,187	-8.9%
Total show less	37,337	36,700	-637	-1.7%
User-days	272,775	270,174	-2,601	-1.0%
User-day liked %	66.45%
[65.80%, 66.54%]	66.86%
[66.34%, 66.94%]	+0.41%	+0.6%
[+0.0%, +1.4%]
User-day show more %	0.75%
[0.68%, 0.79%]	0.80%
[0.72%, 0.85%]	+0.05%	+6.7%
[-2.8%, +17.4%]
User-day show less %	2.85%
[2.68%, 2.91%]	2.87%
[2.71%, 2.91%]	+0.02%	+0.7%
[-5.1%, +6.0%]
Requests per user-day	20.52
[19.7436, 20.4654]	20.94
[20.1519, 20.8653]	+0.42	+2.1%
[-0.6%, +4.7%]
Likes per user-day	12.11
[11.5858, 12.1550]	12.69
[12.1086, 12.7420]	+0.57	+4.7%
[+1.0%, +8.3%]
Show more per user-day	0.0491
[0.0336, 0.0625]	0.0452
[0.0357, 0.0527]	-0.0039	-8.0%
[-39.6%, +23.5%]
Show less per user-day	0.1369
[0.1166, 0.1516]	0.1358
[0.1174, 0.1487]	-0.0010	-0.8%
[-19.2%, +17.6%]
11165
Reposted by angela zhou
Christopher Whitaker @civicwhitaker.com · 13/09/2026
If the Speaker feels they don’t have enough expertise there’s a way to bring it back - The Office of Technology Assessment! www.liberalcurrents.com/congress-nee...
liberalcurrents.com
Congress Needs Its Tech Experts Back
If Congress is to tackle issues from DOGE sabotage to AI to social media, it needs to resurrect its Office of Technology Assessment.
1331183
angela zhou @angelamczhou.bsky.social · 12/09/2026
waking up to - **gestures wildly** all of this - dario's letter, sama's tweet. I think it's impt now that folks understand why voluntary regulation with self-chosen auditors is not good.
1324
Reposted by angela zhou
Ruo Shui @ruoshuiresearch.bsky.social · 12/09/2026
just in case anyone wanted a table for regulatory capture
0175
Reposted by angela zhou
Jeremy Sheff @jnsheff.bsky.social · 12/09/2026
The first element of this strategy (external private monitors against systemic social risk) was tried in the aughts with credit rating agencies in the mortgage-backed securities market. It failed. darioamodei.com/post/we-must...
darioamodei.com
Dario Amodei — We Must Pace the Frontier
111641
Reposted by angela zhou
Ben Recht @beenwrekt.bsky.social · 12/09/2026
Guess what, folks, METR is not a third party.
Embedded Evaluators. Each frontier AI company commits to giving ongoing, employee-like access to a team of embedded third-party evaluators (such as METR), whose role is to verify adherence to safety practices and commitments, report incidents, and help assess the alignment of not just completed AI models but training pipelines and processes. This is the key step for verifiability of any pacing commitments, and has precedent in the banking industry, which sometimes involves regulatory “supervisors” embedded along with employees. Anthropic is unilaterally committing to this step now. We intend this to be part of a broader push to redouble efforts on our safety and alignment work.
2706
Reposted by angela zhou
Neil Shephard @neilshephard.bsky.social · 11/09/2026
Been looking at Richard Samworth and Rajen Shah's new CUP textbook on modern statistical methods. Really nice selection of topics and pace. www.cambridge.org/core/books/m...
095
angela zhou @angelamczhou.bsky.social · 11/09/2026
✊
0103
Reposted by angela zhou
Juspreet Singh Sandhu @freegaussians.bsky.social · 11/09/2026
Sticking to some good-old math: arxiv.org/abs/2609.08169 This came out of the PHA series of papers as @jtnshi & I observed one could Guerra interpolate the planted-SK with SL field to CWRF. We were curious to develop a sharp understanding of the overlaps under the Gibbs at high-temp for the CWRF.
arxiv.org
On overlap concentration in the Curie-Weiss Random Field model
We show that the overlaps of two independent replicas drawn from the Gibbs measure of the Curie--Weiss model with random field (CWRF) exhibit sub-Gaussian tails when the inverse temperature $ β< 1$ an...
242
angela zhou @angelamczhou.bsky.social · 10/09/2026
Big step for auditable AI in California: Newsom signed SB 813, creating infrastructure for credible independent AI verification. Now we need researchers involved in making that verification meaningful - and eventually mandatory.
071
angela zhou @angelamczhou.bsky.social · 09/09/2026
blog: I argue that we need auditable AI now, because current AI developments threaten auditability of AI systems for ordinary regulatory compliance, and we can't figure out “what to do next about AI” without common knowledge about AI (mis)alignment. separatedhyperplane.substack.com/p/we-need-au...
separatedhyperplane.substack.com
We need Auditable AI to figure out “what to do next about AI”
Mandatory, not voluntary, third-party governance, accountability, and transparency
1306
Reposted by angela zhou
Naomi Saphra @nsaphra.bsky.social · 08/09/2026
Interestingly, there are not one but two breakthroughs from openly human-led teams related to NS. BOT have released incomplete proofs early to maneuver around the openai "scoop". (The other is from Anima Anandkumar's group and has less drama involved.)
mathstodon.xyz
Terence Tao (@tao@mathstodon.xyz)
By sheer coincidence, another completely independent result on the Euler blowup question has just been released by Ganeshram, Duruisseaux, and Anandkumar https://anima-ai.org/2026/09/07/stable-singula...
1224
Reposted by angela zhou
A. Feder Cooper @afedercooper.bsky.social · 06/09/2026
Prospective copyright plaintiffs have started asking me my opinion about citing "Alignment Whack-a-Mole" in litigation. It reports that fine-tuning makes frontier LLMs reproduce up to 85-90% of copyrighted books. I don't think the headline results hold up: afedercooper.info/whack-a-mole/
afedercooper.info
Playing Whack-a-Mole with misconceptions about memorization, extraction, and copyright
A response to Alignment Whack-a-Mole: why its headline book-memorization coverage numbers rest on a measurement procedure that can't separate memorization from coincidence or prompt leakage.
12112
angela zhou @angelamczhou.bsky.social · 08/09/2026
You should read this statement. Difficult to follow all the details at times, but (no surprise) it seems that all the corporatization of math with gazillions of $$ at stake has ratcheted up the toxicity in math. And disappointingly, former academics (Seb) as front-gunners for the coporate machine
2245
angela zhou @angelamczhou.bsky.social · 06/09/2026
070
angela zhou @angelamczhou.bsky.social · 05/09/2026
if you need me, you'll find me at the joint misery of: * the federal govt is destroying most things i care about, every day, for my research on social services * there is no accountability for frontier AI & my SV-friends' ability to buy multi-million dollar homes is reliant on this continuing
1251
angela zhou @angelamczhou.bsky.social · 04/09/2026
So ... what do all the freaked out / guilty-feeling AI researchers do to get mandatory incident reporting in place for AI development? is transparency going to be just based on handshake agreements with an in-crowd?
130
angela zhou @angelamczhou.bsky.social · 04/09/2026
www.wikiservice.at/probier/wiki... More message boards? For accessing PUMS ? Texas poverty research ?
wikiservice.at
ProbierWiki: RecentChanges
000
Reposted by angela zhou
lastpositivist.bsky.social @lastpositivist.bsky.social · 04/09/2026
If you want to read the paper that will soon ruin @danielmalinsky.bsky.social's life you can check it out here. Forthcoming in Journal of Causal Inference. www.liamkofibright.com/uploads/4/8/...
liamkofibright.com
76719
Reposted by angela zhou
professor tools for commensality (red scare edition) @inquiline.myatproto.social · 01/09/2026
boosting for infrastructure sickos
0141
Reposted by angela zhou
professor tools for commensality (red scare edition) @inquiline.myatproto.social · 01/09/2026
for additional truck WTAF, may i recommend this frontline on how the trucking industry has successfully blocked important safety regulation since this report was made www.pbs.org/wgbh/frontli...
2124
Reposted by angela zhou
Katie Bergh @katiebergh.bsky.social · 01/09/2026
Texas' SNAP application backlog grew from 3,700 in January to more than 150,000 in August as the state implemented major changes in response to the Republican reconciliation law's unprecedented SNAP cuts.
11914