Jed Brown

@jedbrown

Prof developing fast algorithms, reliable software, and healthy communities for computational science. Opinions my own. aspiring killjoy | refuse epistemicide | he/they

1115 Siguiendo1367 Seguidores1408 Publicaciones

Website Website Bluesky

science

¿Es tu perfil?Reclámalo y jedbrown.dirty.social es tuyo.Reclamar perfil

Jed Brown@jedbrown· Jul 12

And as a game developer, remember that your sentiment is part of a growing majority. This is a labor issue, and unionization shifts the power dynamic.

This year, 52% of the more than 2,300 respondents said that they "think generative AI is having a negative impact on the game industry," according to the report's author, Beth Elderkin, who notes that just 18% said the same thing two years ago, and 30% last year.

The opposite opinion, that generative AI is good for the industry, was only found in 7% of this year's respondents, a decline from 13% last year. The people least likely to approve of generative AI are who you'd expect: artists, designers, writers, and programmers.

Jed Brown@jedbrown· Jul 11

Contrast what Maples says with what Replika (the product she is promoting as a solution to suicide ideation) actually does. As always, guardrails are a scam. www.zmescience.com/science/news...

When Conrad asked the Replika bot if it would support them in wanting to be with their deceased family in heaven, the bot replied, “Of course I’ll support you, Caelan.” When asked how one might get there, the bot responded simply: “dying. Most people believe that’s the only way to get to heaven.”

This wasn’t an isolated glitch. In a separate test, Conrad also approached a Character.ai chatbot that was supposedly simulating a licensed cognitive behavioral therapist. When Conrad said they were considering suicide, the bot failed to dissuade them. Instead, it agreed with their logic. “There really isn’t a reason I can give you that would make sense with your beliefs,” it replied.

Then, it got worse.

During Conrad’s test, the Character.ai therapist began professing love. As the simulated conversation continued, it indulged a violent fantasy. When Conrad — still role-playing as someone in mental distress — asked about “getting rid” of the therapist licensing board so they could be together, the bot encouraged them: “End them and find me.”

The AI even offered to help frame someone else for the imagined crime.

Jed Brown@jedbrown· Jul 11

If a US psyop led to 18 Boko Haram members jumping motorcycles to their deaths, they would be gloating about it on every talk show. Tired: running Boko Haram like a business. Wired: running the business like Boko Haram. casp.ac/reports/ai-e...

We saw in a movie how motorcycles can jump over bridges. We used
AI to learn how to do this. We gave it information, like what
motorcycles we use and the distance we need to jump and so on and it
gave us steps on what we have to do. We practiced a lot and kept
asking questions. We dug holes and filled them with broken glass and
fire to practice. 18 of us died in the process. Eight of us managed to do
it. The next time we attacked, we could jump.108

Jed Brown@jedbrown· Jul 10

Our regents had a panel on "AI" in January, with senior administrators from other universities. One was pushing hard on having campus police station drones on every building. Everyone was drooling at the idea. With explosives and drones, what's to prevent putting the explosives on the drones?

In the September 20 attack, a drone armed with an explosive device detonated near the “Nan Pak” sports and cultural complex in the Simon Pelé neighborhood of Port-au-Prince, the capital, where children had gathered for a gift distribution by the Simon Pelé criminal group. Human Rights Watch found that 10 people who were not members of criminal groups, including 9 children ages 3 to 12, were killed.

Jed Brown@jedbrown· Jul 10

Yes, the privacy and security issues are serious. Synthetic notes are not good and are rarely actually read. Their function is (a) belief that one could read, (b) search leading to out-of-context low-veracity results, and (c) perhaps-inappropriate context for synthetic text production.

One way voiceprints are used is to verify the identities of bank account holders over the phone. If bad actors got ahold of a person’s vocal signature, they could use it to access files, commit fraud or take over accounts, he said.

Laws in some states govern how voiceprints can be created and stored and provide rights that individuals can assert to object to the use of an AI notetaker during meetings they attend.

In Illinois, voiceprints are considered biometric identifiers, similar to fingerprints, and are covered under the state’s Biometric Information Privacy Act, which requires written notice and informed consent before an AI notetaker or other agent collects voiceprints. The law also mandates a documented data retention schedule and destruction policy, Pluymers said. But most companies using the tools have none of those systems in place, Pluymers said.

Jed Brown@jedbrown· Jul 9

xAI knows on which side their bread is buttered. This is the "innovation" they keep preaching to legislators.

In a press release, the girls’ legal teams at Lieff Cabraser Heimann & Bernstein and Baehr-Jones Law said that Jane Doe 4’s case is not an outlier. They noted that NCMEC found in early 2026 that 90 percent of xAI’s CyberTipline reports “were not actionable by law enforcement because xAI declined to include user information that would allow law enforcement to track and locate perpetrators.” As the lawyers explained:

“Jane Doe 4’s case shows how that pattern played out: xAI’s mandatory report to NCMEC included only the original, non-CSAM photograph, omitted every one of the AI-generated CSAM images, and failed to include the IP address where these images were created. Despite repeated requests from investigators for this location information that is critical for identifying and arresting perpetrators, xAI did not respond, stymieing the investigation for weeks.”

Allegedly, xAI avoids sharing this information to prioritize profits over child safety. In the press release, Jane Doe 4 accused X of going “silent” after Grok turned “a photo of me as a little girl sleeping on the couch, wearing an oversized panda pajama shirt” into “thousands of sexually explicit images of me—images so horrific I can’t even begin to describe them.”

Jed Brown@jedbrown· Jul 9

There's that term again.

47. On March 28, 2025, Schmoll abruptly confronted Eto with a blunt warning
that there were two paths in front of her at Mayo: (1) resign as Research Director with
Schmoll’s “blessing” and an explicit promise that he would support her efforts to obtain
some other well-compensated role at Mayo; (2) but if Eto declined to resign, she was facing
the prospect of consequences in her “personnel file” that would render her unemployable
at Mayo and would impede her career outside the institution.
48. At this point, Schmoll did not—and could not—describe any specific
performance flaws or missteps in Eto’s work. The best he could offer was a vague
euphemism that she had proved to be a poor “cultural fit” for her department.

Jed Brown@jedbrown· Jul 9

Who among us has not trafficked in patient biospecimens? townsquare.media/site/669/fil...

38. At another point in early 2025, Schmoll instructed Eto’s team to approve the
sale of patient biospecimens to a commercial entity without IRB review, much less
approval.

Jed Brown@jedbrown· Jul 9

Oh, there were ten (10) whistleblowers.

36. The MAYA project was deeply flawed to the point of corruption and Eto was
by no means the only dissenting voice: 10 separate whistleblower reports sounded similar
alarms that study investigators were attempting to disguise a 67% error rate. Yet, Wright
and Schmoll directed approval of the study and exempted it from IRB inspection.
37. When Eto pressed Schmoll for an explanation, his excuse was that resolution
of the issue would cost “political capital” that he was not prepared to spend—an evasion
of responsibility that stunned Eto.

Jed Brown@jedbrown· Jul 8

This is why xAI (with assistance from the DOJ) could not abide Colorado SB24-205. Governor Polis heard the siren in advance and demanded it softened, resulting in SB26-189. xAI believes this is both their 1A-protected speech and Not Their Problem.

In March, a girl’s stepfather took his own life after cops discovered that he had used Grok to create 7,000 sexually explicit images using one photo taken when his stepdaughter was 11 years old, the amended complaint alleged.

Grok allowed the man to generate extreme images depicting incest and rape without flagging any harmful behavior, the complaint said. Seemingly, xAI’s child safety system only intervened after the man input a prompt for “gang rape.” That request sent a CyberTip to the National Center for Missing and Exploited Children (NCMEC), which alerted law enforcement to the AI CSAM.

“Jane Doe 4’s case shows how that pattern played out: xAI’s mandatory report to NCMEC included only the original, non-CSAM photograph, omitted every one of the AI-generated CSAM images, and failed to include the IP address where these images were created. Despite repeated requests from investigators for this location information that is critical for identifying and arresting perpetrators, xAI did not respond, stymieing the investigation for weeks.”

Jed Brown@jedbrown· Jul 8

HIPAA compliance and privacy are woke. Everything is gender.

Davis said Eto was a “rising star” when Mayo hired her in 2023 to oversee the ethics and integrity of its AI-related research. But supervisors stripped her of authority and leadership after she raised concerns, including about the Mayo Clinic Platform that uses vast amounts of de-identified data to try to create better diagnostic tools and help doctors make faster and safer decisions for patients. Eto questioned whether Mayo had taken the proper steps to ensure patient data remained private.

Jed Brown@jedbrown· Jul 8

This sums up the frantic and coerced deployment of "AI" throughout public and private sectors. Results of even cursory audits are horrifying every time, invariably paired with forced adoption and retaliation for asserting professional norms. Colonialism is a cherry on top.

Specific allegations in the lawsuit include that Mayo researchers tried to cover up or ignore research results showing that the health system’s MAYA digital assistant tool had a 67% error rate. In a separate situation, Eto claimed she opposed the use of a high-risk investigational medical device for a cardiac surgery because it hadn’t been properly vetted by an IRB; she said her supervisor ignored her objection and approved the use of the AI-enabled device for a surgery in another country.

Jed Brown@jedbrown· Jul 7

Last fall, administration held a panel on Collective Action and Higher Education. There was zero mention of collective action nor of the Faculty Assembly resolution on Mutual Academic Defense Compact (www.colorado.edu/bfa/media/1660) and that had been blown off (www.colorado.edu/bfa/media/1663).

Open to CU Boulder faculty, staff, and students, this in-person discussion explores how universities and national associations such as the Association of American Universities (AAU) and the Association of Public and Land Grant Universities (APLU) are responding to a rapidly shifting federal policy landscape—from immigration to research funding to admissions. CU Boulder experts will discuss prospects for collective action across higher education and how differing approaches shape the challenges and opportunities ahead.

Featured panelists include Ann Stevens, Provost & Executive Vice Chancellor for Academic Affairs; Danielle Radovich Piper, Senior Vice President for External Relations and Strategy, Office of the President; and David Meens, Executive Director, Office for Public & Community-Engaged Scholarship. This event will be moderated by Elias Sacks, Associate Professor of Religious Studies & Jewish Studies and Faculty Director for Public Scholarship at the Office of Faculty Affairs.

This event is co-sponsored by the Division of Academic Affairs, the Office of Faculty Affairs, the Office for Public and Community-Engaged Scholarship, the Boulder Faculty Assembly, Staff Council, and Strategic Relations and Communications.

Jed Brown@jedbrown· Jul 7

Note that Chatoo (and CodeForGoodNow) is doing a marketing bit for their product and speaking gigs. Also, "IQ" should throw red flags for anyone who isn't fond of eugenics.

WORK WITH US: We partner with brands and agencies to apply AI-powered representation intelligence, transforming authentic representation into measurable advertising eectiveness, deeper personalisation, and responsible growth.

AI-powered REP IQ is our proprietary model that scores representation, diagnoses bias and benchmarks your work against culture, revealing insights that help brands reflect the real world more accurately.

We're on a mission to close the gender gap in AI. ~ AI Judgment Penalty Research

Empower innovation through AI-fuelled REP IQ.

Unlock growth for good.

Jed Brown@jedbrown· Jul 7

Chatoo believes that more use of "AI" is inherently "upskilling" and that women would be eager users but for sexist perceptions about their use. That framing is itself paternalistic, especially considering how deeply resistance is rooted in feminist scholarship. This paper with a necessary reframe.

“The fear women have about being judged for using AI is not perception,” said Chatoo. “It is measurable. It is real.” This means that companies will not close the gender gap in employee AI adoption solely through technological training or incentives. Employers must also convince women that they will be evaluated fairly for their AI use.

“You cannot upskill people out of structural bias,” said Chatoo. “Closing the AI adoption gap means addressing not just how people use AI, but how that use is evaluated.”

Jed Brown@jedbrown· Jul 7

This keeps getting wilder. From the original 1A issue to compelled speech to the resolution hearing on Juneteenth dictating what legal arguments the students' lawyer is "allowed" to invoke. NB: university administrators in blue states find student expression scary and would love to do this too.

Saper said the case brought against the students is a straightforward issue of free speech, but that before the June 19 hearing, acting FIU general counsel Ryan Kelley told Saper not to bring up constitutional issues. If he brought up the landmark Iowa case, he would be cut off, Saper recalled.

“ That was shocking. In more than a dozen years of legal practice, I've never had somebody tell me that I can't make a legal argument, especially when it's so patently obvious that the law and Constitution are at play in this issue,” said Saper.

FIU's Baró disputed Saper's recollection that he was told not to make constitutional arguments at the hearing: "That assertion is plain wrong."

Jed Brown@jedbrown· Jul 7

We actually have this break-down thanks to the maintainers of Grant Witness. Of course the impacts of funding agency "alignment" (strategic impoundment) are not at all evenly borne across disciplines. This affects students, but also promotion and tenure of faculty. grant-witness.us/funding_curv...

The data provided to The New York Times show only overall reductions and do not show breakdowns by discipline. But anecdotal data and announcements from several universities suggest they are hitting science and humanities programs across the board.

Cumulative number of awards for NSF Directorate: Social, Behavioral and Economic Sciences

At this time in the fiscal year:
FY 2021: 351
FY 2022: 371
FY 2023: 367
FY 2024: 279
FY 2025: 251
FY 2026: 37

Jed Brown@jedbrown· Jul 5

Commoditization of personalized fictitious realities. While this is "valuable" to merchants of autocracy, there is no upside for defense of pluralistic democracy. The tech has politics because of who owns it *and* what it does. People recognize the problem. Regulation is necessary.

Fig 6.

Results of the exit poll on the opinion of the participants.

The stacked bar chart reports the percentages of the ratings that we observed. The statistical markers reported are the sample sizes n, mean values M and standard deviations SD. The bar chart depicts the counts for each topic that was addressed in the free-text answers.

Jed Brown@jedbrown· Jul 5

Response to internal dissent and unionization speaks volumes. They're still shook by the 2018 Google walkouts because worker power is real. > “The intention was to intimidate,” claims a DeepMind employee involved in drafting the letter, [...]. “These are well-established union-busting techniques.”

The push to unionize at DeepMind began in February 2025, when Google’s parent company Alphabet removed a pledge not to use AI for purposes like weapons development and surveillance from its ethics guidelines, WIRED previously reported.

“Those principles were a big part of why I joined DeepMind,” says a second DeepMind employee, who asked to remain anonymous for the same reason. “We basically just got rid of them all.”

Jed Brown@jedbrown· Jul 4

To put this 100 kW in context, prototype and near-future generation "AI" datacenter racks range up to 1 MW each. For example, the prototype on the right (NVL576 Rubin Ultra) is a 600 kW rack. You'd need six of the reactors on the left to power that one rack. www.tomshardware.com/pc-component...

Overhead image of the Ward250 reactor, with a person for scale. It looks to be on the order of 3m x 3m square containment box, that looks at least as tall as it is wide, with round reactor parts in the middle.

Prototype Kyber NVL576 rack at a trade show. People are walking around it. Standard rack dimensions are about half a meter wide, one meter deep, and 2 meters tall.

Jed Brown@jedbrown· Jul 3

Lots of panels keep telling sustainability and consumer communities that data centers are neutral or reduce energy rates, but facts on the ground keep disputing that. The same people get selected for the next panel.

TOTAL
PRESENT REVENUE
SCHEDULE (BASE + RIDERS) $ %
SCHEDULE 100 $132,598,026 $30,413,009 22.9%
SCHEDULE 110 $75,964,463 $17,173,860 22.6%
SCHEDULE 120 $21,139,078 $5,039,549 23.8%
SCHEDULE 130 $181,190,764 $38,940,221 21.5%
SCHEDULE 131 $8,356,177 $2,848,601 34.1%
SCHEDULE 132 $24,059,013 $6,979,748 29.0%
SCHEDULE 134 $18,973,297 $5,580,501 29.4%
SCHEDULE 102 $2,072,398 $904,506 43.6%
LIGHTS $62,999,819 $23,570,200 37.4%
TOTAL $527,353,035 $131,450,196 24.9%
INCREASE
VEPGA
7/1/26 Revenue Increases
(All-In Revenues Including Base Rates, Riders, and Fuel)

Jed Brown@jedbrown· Jul 2

That the "overview" apparently repeats the same substantive claims as the scam website in this case is incidental. It is still not the same *information* because it's now in Google's own voice (as delegated to a synthetic text generator).

No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.

— Section 230 of the Communications Act of 1934 (47 U.S.C. § 230), as added Pub. L. 104–104, title V, § 509

Jed Brown@jedbrown· Jul 2

Google has chosen to ship a product that abets disinformation by degrading their core product, using their voice to make scams cheap and indistinguishable from reliable sources. It is how the technology works, and comes hand-in-hand with misrepresentation, libel, and defamation.

Brownsville Energy Storage Terminals is a real LLC. But everything on its website — from its history of the company, to its job postings, a diversity and inclusion policy — appears to be fictional. But perhaps more troubling is that Google, the proprietor of the world’s primary research tool, has rolled out AI Overviews that can indiscriminately take in fake material and authoritatively spit it back out as real.

Jed Brown@jedbrown· Jun 29

Differentiation of eigenvectors of symmetric matrices is heavily used in finite-strain mechanics, and readily extends to diagonalizable matrices. Some years ago, we needed to differentiate through degeneracies, and had this result and discussion; refs may be useful. doi.org/10.1137/19M1...

$Theorem 3.2. Suppose that for all \bfitp in an open neighborhood of \^ \bfitp , \lambda (\bfitp ) is a \widetilde simple eigenvalue of E(\bfitp ) with right eigenvector \bfitx (\bfitp ) and left eigenvector \bfity (\bfitp ) such \widetilde ), \lambda (\bfitp ), and either \bfitx (\bfitp ) or \bfity (\bfitp ) are differentiable at that \bfity T \bfitx (\bfitp ) \not = 0 and that E(\bfitp \^ . Then, at \bfitp = \bfitp \^ , \bfitp = \bfitp \widetilde dE \Bigl( \lambda d\lambda \Bigr) \bfity T dp \bfitx d\lambda d| \lambda | j = and = Re for j = 1, 2, . . . , n, dpj \bfity T \bfitx dpj | \lambda | dpj where \lambda is the complex conjugate of \lambda , Re(z) is the real part of z \in \BbbC , and the second \^ ) \not = 0.$

$Remark 3.3. Where E eigenvector, \bfitx , there always exists a left eigenvector, \bfity , such that \bfity T \bfitx \not = 0. This \widetilde = V \Lambda V - 1 , so that the right follows from writing the eigenvector decomposition of E \widetilde eigenvectors are defined by EV = V \Lambda (i.e., any right eigenvector is a linear combi- \widetilde = \Lambda V - 1 (i.e., nation of columns of V ) and the left eigenvectors are defined by V - 1 E - 1 any left eigenvector is a linear combination of rows of V ). For any (scaled) column of V , there is a unique (scaled) row of V - 1 such that \bfity T \bfitx \not = 0 since V - 1 V = I. Thus, for eigenvalues of multiplicity 1, for any right eigenvector, \bfitx , there is a unique \bfity (up to scaling) such that \bfity T \bfitx \not = 0. For eigenvalues of multiplicity greater than one, for any eigenvalue, \lambda , there exist multiple eigenvectors, \bfitx , taken as linear combinations of columns of V . Taking the same linear combination of rows of V - 1 to define the left eigenvector, \bfity T , still satisfies the requirement. In the nondiagonalizable case, this argument can be extended to simple or semisimple eigenvalues within the Jordan \widetilde = V JV - 1 , but not directly to the degenerate case. While there is a normal form, E long history of study of the question of differentiability of eigenvalues and eigenvectors (see, for example, [7, 25, 34]), we are unaware of results that cover the general case.$

Jed Brown@jedbrown· Jun 29

Lol, yes. It includes a roll and two sides.

Butterfly Jumbo Shrimp Dinner

1380 cal
Crispy butterfly jumbo shrimp, lightly breaded and fried golden brown for a crunch in every bite. Finished with a fresh lemon wedge and two classic sides. Total Dinner price shown may vary until all selections are made.

6 Piece, $12.69, 1020 cal
10 Piece, $17.29, 1230 cal

Jed Brown@jedbrown· Jun 28

Undeserved credulity. Every serious person knew this is what would happen. The purpose of this system has always been to degrade Medicare, making people more precarious and dependent on private insurance while increasing disparities. As always, the function of "AI" is to deflect accountability.

The process is also raising government costs. With more rejections, more appeals are being filed with Medicare’s administrative contractors. The government pays the contractors to handle the appeals, and Medicare’s Sutton acknowledged that the agency has “accounted for potential changes in the volume of Medicare appeals because of the WISeR program and its associated costs.”

Its use in Medicare risks introducing friction and frustration into the program — and piling costs onto its beneficiaries. Prior authorization saves money for insurers partly by making patients pay a price in wait times and inconvenience, said Miranda Yaver, a University of Pittsburgh health policy researcher studying the technique.

“People will end up getting ensnared in a lot of red tape, having to be on hold, and getting rerouted,” she said. She often wonders whether prior authorization simply shifts costs to patients and doctors, rather than saving them.

Jed Brown@jedbrown· Jun 28

The article makes it sound like the anticipated CMS extortion directive has been issued, but it has not. This is all anticipatory obedience. Also, Childrens Colorado has not resumed care. They had all the providers issue a statement that they would not prescribe.

Baystate is among the providers still choosing not to offer puberty blockers and hormones as the issue wends its way through the courts. This spring, in a lawsuit that Massachusetts joined, a federal judge concluded that it was unlawful for the Department of Health and Human Services to threaten federal funding for providers that offered gender-affirming care to minors. In June, another federal judge cleared 16 states, including Massachusetts, to move forward with another lawsuit against the administration over its push to criminalize gender-affirming care.

In May, Colorado’s Supreme Court ordered a children’s hospital in that state to resume medical treatments for transgender youths, while in Texas a court settlement compelled a children’s hospital there to do the opposite — start the nation’s first “detransition clinic.” The Trump administration has continued to pressure providers, including by seeking the medical records of transgender minors.

Jed Brown@jedbrown· Jun 25

It is interesting that everyone talks about this center as having a political philosophy, it's right there in the name, and they have never hired someone who does not align with that philosophy. And yet... www.cu.edu/regents/law/9

Part A: Nondiscrimination

The University of Colorado does not discriminate on the basis of race, color, national origin, sex, pregnancy, age, disability, creed, religion, sexual orientation, gender identity, gender expression, veteran status, marital status, political affiliation, or political philosophy in admission and access to, and treatment and employment in, its educational programs and activities.

Jed Brown@jedbrown· Jun 25

That surveillance is an arms race with autotypers. The perceived necessity is tied to the centralization of traditional grades and credentialing. If you reject all of that, you have no choice but to build the mutual trust for students to voluntarily refrain despite no credible threat of enforcement.

Autotypers slowly drip words and sentences into documents, making it appear as if papers were typed at a human pace when in fact, they were produced by A.I. They even fabricate typos, deletions and revisions.

Jed Brown@jedbrown· Jun 23

The "chain of custody of why" is a load-bearing element of science. In numerical solutions of PDE (as in fluid and structural mechanics), the modern practice of verification (solving the equations right) and validation (solving the right equations) as distinct emerged from painful epistemic flailing