Ambrose (@ambrosia-engine) · DIRTY

Ambrose

@ambrosia-engine

PhD student in Informatics: AIAI: Automated Reasoning, Agents, Data Intensive Research, Knowledge Management. No, I don't know why the title is so long. Also Affiliate of the Center for Technomoral Futures. Views are my own.

2.598 Folgt1.150 Follower1.331 Posts

Ist das dein Profil?Übernimm es — dann gehört dir ambrosia-engine.dirty.social.Profil übernehmen

Ambrose@ambrosia-engine· 19d

Wait a minute, am I reading this right? I actually could apply for this?

Review of applications will begin on July 31, 2026. If you already possess a doctoral degree at the time of hiring, you will be hired as a Postdoctoral Associate. If you do not (i.e., you will be defending during your first year at CSMAP), you will be initially hired as a Researcher and then will be transitioned to a Postdoctoral Associate upon completion.

Ambrose@ambrosia-engine· 19d

🫩 In the words of a billionaire "I'm the problem, it's me." 🫩

Eugenics Record Office
Eugenics research institute at Cold Spring Harbor, New York

The Eugenics Record Office (ERO), located in Cold Spring Harbor, New York, United States, was a research institute that gathered biological and social information about the American population, serving as a center for eugenics and human heredity research from 1910 to 1939. It was established by the Carnegie Institution of Washington (Carnegie Science), Station for Experimental Evolution, and subsequently administered by its Department of Genetics.[1]

Both its founder, Charles Benedict Davenport, and its director, Harry H. Laughlin, were major contributors to the field of eugenics in the United States. Its mission was to collect substantial information on the ancestry of the American population, to produce literature fueling the eugenics movement, and to promote the idea of race betterment.

Ambrose@ambrosia-engine· Jul 9

Checks out.

"Ethics declarations
Competing interests
Two authors of this article, SB and TF, are involved in organising the European Conference on Ethics and Integrity in Academia 2023, co-organised by the European Network for Academic Integrity. This conference receives sponsorship from Turnitin and Compilatio. This did not influence the research presented in the paper in any phase."

Turnitin and Compilatio are underlined

Ambrose@ambrosia-engine· Jul 7

Types to look out for, from: doi.org/10.1016/j.ip...

Table:

Table 1. Malicious social media bots types.
Type name Brief description
Cashtag piggybacking bots “Coordinated groups of bots that promote low-value stocks by exploiting the popularity of high-value ones” Cresci et al. (2019a).
Spam bots “Bots that distribute malicious links, unsolicited messages and hijack trending topics on the social network” Wang (2010).
Astroturfing bot “Create the appearance of widespread support for a candidate or opinion” Ratkiewicz et al. (2011).
Sybils “Pseudonymous identities, i.e. user accounts, used for a disproportionately large influence” AlRubaian, Al-Qurishi, Rahman, and Alamri (2015).
Fake accounts used for botnet Command & Control “Fake accounts that share posts which includes encrypted commands for a botnet attack. They use the OSN platform as a C&C network” Sebastian, Ayyappan, and Vinod (2014).
Pay bots “Illicitly make money. For instance, by copying post content from a respected sources but paste in micro-URLs that direct users to sites that pay the botmaster for directing traffic to the site” Subrahmanian et al. (2016).
Social botnets in political conflicts “Bots involved in political discussions. They Spam the discussions with biased contentions or unrelated content to deflect the readers” Abokhodair et al. (2015).
Infiltration of an organization “Bots that act as social network friends in order to gain potentially harmful information from users” Elyashar, Fire, Kagan, and Elovici (2016).
Influence bots “Realistic automated identities that illicitly shape discussion” Subrahmanian et al. (2016).
Doppelgänger bots “Clone the profiles of ordinary people on Twitter to create real-looking fake identities and use them in malicious activities” Goga, Venkatadri, and Gummadi (2015).

Ambrose@ambrosia-engine· Jun 29

Love a good diagram.

"That ecosystem can reach both voters and politicians in extremely powerful ways. Sufficiently powerful to persuade any of us. It doesn’t matter how smart you are, if you’re surrounded by a personally-tailored wall of disinformation, you are vulnerable."

This precedes a very sick diagram showing how political influence shapes narratives we are exposed to.

From left to right in blue the boxes show right wing influencers, media, lobbying groups, to major donors. Thus is above unboxed examples from left to right these are Tate / Robinson, social and mainstream media, think tanks, traditional lobbying, and political parties. At the bottom arrows go from the right wing influence and media to UK population. There is a label with an arrow coming in from the right labeled as lived experience.

The diagram suggests political extremists, state actors, and big tech / big oil / big etc spend money to lobby and donate to political parties, media, and right wing influencers. The only oppositional source of information is lived experience.

Ambrose@ambrosia-engine· Jun 28

This review really goes for the jugular.

"Any serious moral accounting of the Weather Underground might consider how they would be remembered had they stormed a federal facility and indiscriminately fired 2,000 rounds of live ammunition, murdering 43 people in under 20 minutes (a crime actually committed by the New York State Police and correctional officers when they retook Attica in 1971). How might we judge them had they opened fire on unarmed protestors, injuring 28 and killing three by shooting them in the back (a crime actually committed by state police and highway patrol officers in the 1968 Orangeburg Massacre)? How might we remember the Weather Underground had it been J. Edgar Hoover and not Fred Hampton who lay in a drugged haze as armed men stormed into his home and shot him in his sleep?"

Ambrose@ambrosia-engine· Jun 28

Just fantastic requirements layed out here. Great work from the many orgs and researchers who wrote this report and these recommendations. We all should be emailing our MPs to see if they actually care enough to regulate big tech by actually enforcing the OSA with this guidance.

TRANSPARENCY
a. Services must have a comprehensive and proactive strategy for providing
stakeholders with clear and accessible information about the design,
operation and remaining risk present on the service. this must include
the risks identified through ongoing risk assessments, including those
arising from features, functionalities, and business models, and details of
the specific mitigations implemented, and how the effectiveness of those
mitigations are measured. It must also include clearly assigned points of
accountability to the public and to regulators.
b. Services should publish regular, clear transparency reports that allow
regulators, researchers, and the public to understand how the service
manages risk. these reports should show how much harmful or illegal content
is detected and acted on, what types of enforcement actions are taken and
how quickly after being made aware of the harm, service response to user
reporting and how quickly after a report is received, and how any automated
safety tools are working. they should also report on effectiveness, including
accuracy, errors, and whether risks are being reduced over time. additionally,
reports should cover user complaints, reports, and appeals (e.g. how many
there are, how they are resolved, and how long responses take). the metrics
and measures included in this reporting must be appropriate to deliver the
full picture of safety, and must not be presented in misleading or manipulative
ways.
c. Services should organise information into clear, meaningful categories, such
as the type of risk, the feature involved, and the affected user group. they
should highlight any systemic risks, emerging harms, and the steps taken in

Ambrose@ambrosia-engine· Jun 28

It still suggests age verification, but provides very good requirements for the deployment and outcome based re-evaluation of such systems. My concern here is the security of such sensitive data in the hands of 3rd party companies, but it sounds like a breach would be a change to the risk profile.

all age assurance methods involve personal data processing and must
comply with applicable data protection law. Services must embed privacy
preservation by design and default, minimise data collection to that which is
strictly necessary, and implement safeguards against misuse, discrimination,
and unwanted bias. Systems must not be repurposed for incompatible uses
such as marketing or profiling unrelated to safety outcomes, and must be
designed and operated in full compliance with uK gdPr, the age-appropriate
design code, ISo/IEc 27566-1 (which provides a general framework for age
assurance deployments), and IEEE 2089.1-2024 (which goes into further
detail on achieving age assurance that respects children’s rights).

Services must evaluate the effectiveness of age assurance measures based
on outcomes, not solely on the technical selection of methods. this includes
measuring whether the system correctly identifies under-age users, correctly
identifies adult users, prevents inappropriate access, and reduces identified
risks in practice. Performance must be documented publicly against clear
criteria, including accuracy, privacy preservation, robustness, reliability,
fairness, and user experience, and reviewed regularly.

age assurance systems must be tested to avoid discriminatory outcomes
for users with protected characteristics, including race, disability or socio-
economic background. children and relevant user groups should be engaged
in testing and validation to ensure systems are understandable, accessible,
and effective. alternative methods must be available for users unable to
complete a given assurance method, and transparent challenge and redress
process must be maintained. Ideally, multiple alternatives should be offered
to allow users unable to complete a given assurance method with a choice
of options. for example, for users unable to complete facial age estimation,
their only other option should not be to upload a hard Id document where this
can be avoided, as this presents outsize risks (e.g. to exclusion, privacy and so
on) to groups that are likely already vulnerable (including children)

age assurance obligations include an ongoing duty to review and update
systems, thresholds, and methods in light of emerging technologies, changes
in risk profiles, operational experience or new regulatory guidance. Services
must implement governance processes to ensure age assurance remains
aligned with risk outcomes, user needs, and rights and does not degrade over
time.
h. Minimum criteria includes:
i. Highly effective
ii. Must offer more than one form of age assurance
iii. Must be in full compliance with GDPR, AADC and ICO’s opinion on age
assurance
iv. Monitored and tested.
v. Must protect user’s rights to privacy, anonymity and access to information

Ambrose@ambrosia-engine· Jun 28

I also love the reports inclusion of examples. It mentions apps like Bluesky and Mastodon which have user tools and functionality that make them safer-by-design. While not every problem has easy techno-legal solutionisms, regulators can't just throw their hands in the air and say "too hard."

EXAMPLE
Chelsea Peterson-Salahuddin describes how hate speech content moderation could
be done with reparation in mind, which focuses on “redress, not removal” - redress
which could include apologies from the poster of hateful content, affirmative
actions, and/or wider community development schemes involving marginalised
users to identify further ways that harm could be mitigated and repaired.52
Danielle Cole similarly describes a Black feminist model of social media, in which a
platform would facilitate time-outs and ‘calling in’ of users who have shared hateful
content, and accountability for behaviour would be shared by the online community,
focusing on repair rather than punishment

Ambrose@ambrosia-engine· Jun 28

Lots of attention to dark patterns and persuasive design. Most chatbot products, particularly the ones based upon foundation models, would not meet these standards. Do recall Microsoft's explicit strategy for chatbots is "Make people addicted." jgcarpenter.com/blog.html?bl...

PROHIBITED PRACTICES
Prohibited high risk practices: There are some features and functionalities which should
not be employed at all. Services must not deploy interfaces that manipulate, pressure, or
mislead users into making decisions that reduce their safety, privacy, or wellbeing. This
prohibition applies to all users, with enhanced protections required where children are
likely to access the service.

Prohibition of ‘dark patterns’ for all users: Services must not deploy interfaces that
manipulate, pressure, or mislead users into making decisions that reduce their safety,
privacy, or wellbeing. This prohibition applies to all users, with enhanced protections
required where children are likely to access the service.

Persuasive design here refers to design choices and functionalities that aim to engage
the user, tapping into psychological vulnerabilities all humans have. They utilise the Fogg
principles: linking motivation, ability, and triggers at the right moment to drive behaviour.
Persuasive design is different from “normal” user interface design, which seeks to make
the service easy to use, because of this intentional use of psychological triggers. While we
refer here to specific features, there is a close link between persuasive design and dark
patterns. For the code, the key point is that persuasive design is about those features
that promote repeat or extended use - in non-medical parlance, those features tending
towards addiction. We refer to Ofcom’s commissioned report on children’s financial harm
and persuasive design40. They identify different sorts of features (a classification which
could also apply to dark patterns):
• Risk-based: These included mystery rewards, including paid-for rewards, and
platforms that visually resembled gambling sites;
• Dissociative: Make it harder for users to stay conscious of what and how much
they are spending;
• Misleading;
• Impulsive: Encourage users to make quick, impulsive decisions surrounding
spending;
• Social influence: Exploit and amplify peer pressures such as sharp differences in
default and paid-for cosmetic items – making the paid-for versions feel essential.

Ambrose@ambrosia-engine· Jun 28

These are also great! When designing any tech product we should be taking into account which values are prioritized and traded-off in the design, and how real people are harmed by those choices. Profit is never listed as a company value, but it's almost always the one optimized for in design.

Understanding hazards and mitigation
a. In support of the risk assessment required by the oSa31, services should
generate risk acceptance criteria, taking into account ofcom’s risk register
and relevant risk profiles and taking into account the potential impact on
different groups of users. the criteria should reflect the service provider’s
publicly stated values. Services should develop appropriate metrics for
quantification of risk and develop a baseline standard of hazards and risk of
harm to inform the risk assessment. Services must engage with civil society,
experts, and individuals with lived experience on this process.
b. following testing, and carrying out risk assessments in accordance with
ofcom guidance, document mitigation strategies taking into account the
hierarchy of control: Services must prioritise risk mitigation measures
according to the following hierarchy. (Services must demonstrate that higher-
level controls have been considered before relying on lower-level controls,
and be able to justify why those higher-level controls were not usable)32.
• Elimination: Remove features or system designs that create a foreseeable risk of
harm.
• Substitution or design modification: Redesign features to reduce the likelihood
or severity of harm.
• Engineering and technical safeguards: Implement robust technical systems that
limit or detect harmful behaviour.
• Administrative controls: Actionable and upheld policies, moderation systems and
governance processes to manage risks.
• User-level tools: Mechanisms enabling users to control their experience or
respond to harm.
That the design and operation of the service meets minimum standards of human rights
and non-discrimination, and if it fails to do so, the service or functionality cease to be
deployed until remedial action is taken
Mitigation strategy should be kept updated and respond to any findings developed under
product testing or market surveillance mechanisms.

Ambrose@ambrosia-engine· Jun 28

The independent and transparent oversight mechanisms are critical, we shouldn't let big tech decide if the products they want to sell us are safe and take their word for it. Nor should we just take Ofcom's word for it.

Services must implement governance structures that ensure safety considerations
are embedded in strategic decision-making and product development. this must
include:
a. Board-level oversight of safety risks and named responsible person(s), where
the level of oversight is proportionate to the level of safety risk
b. clear and public allocation of responsibility for compliance with this code
including by specifying a named person with overall responsibility for safety
by design, as well as senior managers in relevant roles
c. regular internal review of safety performance and risk mitigation
measures, conducted by personnel with sufficient resources, expertise and
independence as necessary, and which engages directly with users most
affected by risks
d. conduct rules for senior managers, requiring actions that promote and uphold
user safety, ensure transparency and disclosure, and necessitate due regard
to the best interests of the child
e. Integration of safety-by-design principles throughout the product lifecycle
f. Independent oversight mechanisms

Accountability: Evidence of compliance, documented risk assessment and governance
processes, mitigation strategies, details of board oversight and senior manager
responsible for this documentation. Publicly available documents demonstrating
appropriate consideration (and where feasible mitigation) of potential risks to children

Ambrose@ambrosia-engine· Jun 26

I think it's so interesting in Futurama how much of the show centered around tv. In the year 3000, they're still watching the tv every day together. The show moves a bit away from this later on. I wonder what it means that we consume so much more media in isolation.

Screenshot from Futurama, Season 6, ep 3 - Attack of the Killer App. Pictured: The character Mom in blue. She's clasping her hands, evil-like, and ruminating: "Thanks to twitcher, morons voluntarily spew out every fact I need to exploit them." At this point in the show, the cartoonishly evil tycoon had already been portrayed as using their spectacular wealth and influence to commit many crimes against humanity - among others, inspired by profit or whimsy.

Ambrose@ambrosia-engine· Jun 7

Not this being one of three books I'm flipping through and flopping between today. At least I'm being 360 no-scoped on this book rather than some of his other work 😅

Excerpt from Umberto Eco's "How to write a thesis" - English translation.
"Naturally, the approach depends on the researcher's psychological structure. There are monochronic people and polychronic people. The monochronic succeed only if they work on one endeavor at a time. They cannot read while listening to music; they cannot interrupt a novel to begin another without losing the thread; at their worst, they are unable to have a conversation while they shave or put on their makeup. The polychronic are the exact opposite. They succeed only if they cultivate many interests simultaneously; if they dedicate themselves to only venture, they fall prey to boredom. The monochronic are more methodical but often have little imagination. The polychronic seem more creative, but they often messy and fickle. In the end, if you explore the biographies of great thinkers and writers, you will find that there were both polychronic and monochronic among them."

Ambrose@ambrosia-engine· Jun 5

Oh that's fun, actually reading this exact book right now on recommendation from my supervisor.

Recent studies indicate that hyperreading not only
requires different reading strategies than close reading but also may be involved with
changes in brain architecture that makes close reading more difficult to achieve.
Much of this evidence is summarized by Nicholas Carr in The Shallows: What the
Internet Is Doing to Our Brains. More judicious than Bauerlein, he readily admits
that Web reading has enormously increased the scope of information available, from
global politics to scholarly debates.

Ambrose@ambrosia-engine· Jun 5

Drawing distinctions can be very helpful! Like is there a distinction between defending an atrocity at an unfathomable scale versus defending it at a smaller scale? A distinction between defending a harm that directly impacts millions or a harm that only impacts a small amount of people.

No training energy data are yet available for
GPT-5, but scaling up from GPT-3 and GPT-4
suggests that a next-generation model could
plausibly require on the order of 100 GWh, with
associated footprints of roughly 42,000 tonnes
CO₂e, one billion liters of water, and a land
footprint of 1.5 km² of land, or roughly the size
of 215 football fields. Offsetting this carbon
would require about 700,000 tree seedlings
grown for 10 years, equivalent to roughly the
number of trees in 180 Hyde Parks in London.
This level of electricity use is equivalent to the
annual residential electricity consumption of
approximately 770,000 people in Sub-Saharan
Africa and the water footprint is sufficient
to meet the minimum annual water needs of
roughly 137,000 Sub-Saharan residents.

Ambrose@ambrosia-engine· May 30

I really recommend this book as it has an incredible chapter on this exact topic. Our modern obsession with debate is actually dampening our political reasoning skills.

Cover of "Don't Talk About Politics" by Sarah Stein Lubrano.

Ambrose@ambrosia-engine· May 12

🤢 me when I saw this email

Your overleaf subscription is getting an AI-power-up

Ambrose@ambrosia-engine· May 3

How terrible indeed for advertisers that such a simple tool exists. Will those wicked ne'er-do-wells not think of the starving children of the ad execs? I for one am kept awake at night with worry for the shareholders. How could anyone ever endorse such anti-social behaviour. 😭😭😭

"The average cost-per-click for e-commerce Display Ads is $65.80. So,

$65.80 (CPC) x 2 (number of ads per Ad Nauseam user) = $131.60

216,000 (estimated number of Ad Nauseam users) x $131.60 = $28,425,600.

That’s $28.4 million of invalid clicks in eCommerce, just from Ad Nauseam.

Per day.

It’s a staggering figure. And if you’re an e-commerce PPC advertiser, it’s fairly terrifying."

Ambrose@ambrosia-engine· May 2

This is more important and likely most important to reshaping our society in ways that are desirable. That is building infrastructure, especially infrastructure that allows us to connect with others we otherwise wouldn't and expand our action-possibilities.

Cover of "Don't talk about politics" by Sarah Stein Lubrano. Features the title in black bolded lets, the author in red letters, and a collage of tweets in the background.

Ambrose@ambrosia-engine· May 2

"To these patients, Fraad recommends “not being alone” and embracing “activism, love and solidarity.”"

Similarly, it can’t hurt for us to recognize when we are participating in habits that reflect and exacerbate terra infirma—we can reject predictive betting markets and their janky fake sense of relief, for example, or use tools that strip our feeds of AI slop wherever they find it, demanding a more human internet.

I am trying to acknowledge the political and economic uncertainty and nihilism around me, to live with it and name it. Otherwise, there is always the danger of repression, which leads, according to psychologists, to our splitting into metaphorical parts. The version of myself that tries on tinted sunscreen, makes sure to Docusign contracts, and watches regional UK TV procedurals late into the night co-exists with the version of myself that is hyper-vigilant to the extreme events that keep unfolding.

Ambrose@ambrosia-engine· Apr 25

This isn't rocket science. This techno-legal solutionism isn't going to work and is already not working. Invest in digital literacy for everyone, especially whoever designed these policies which require people to undermine their privacy for access to a still unregulated Internet.

“What started with Australia’s social media ban quickly spread,” Simon commented. “The US and Europe are now catching up, and the scope is already expanding beyond social media.” Early evidence from Australia shows, so far, no clear decline in reported harms such as cyberbullying or image-based abuse.

Besides, recent studies indicate that 61% of Australian children aged 12 to 15 still have access to restricted platforms, while 70 percent say it is easy to bypass the ban, using accounts created by older friends or relatives, or manipulating age-estimation tools, or using VPNs. The country’s own regulator has acknowledged that “a substantial proportion” of minors continue online."

Ambrose@ambrosia-engine· Apr 21

"Sensing a change in the air, Big Tech firms have wasted no time rolling out offerings meant to soothe — and cash in on — Europe’s unease."

The legal intricacies of sanctions aside, this once far-fetched hypothesis is now a widespread concern: A recent survey from SWG and Polling Europe shows that 86 percent of Europeans think a sudden U.S. move to restrict the EU’s access to digital services is “plausible” and “should not be ruled out,” while 59 percent call it “already a real and concrete risk.”

Ambrose@ambrosia-engine· Apr 20

Hearing it is one thing, reading his response, another entirely. It's like Ed Davey was talking to a chatbot.

Transcript of end of the question from Ed Davey to the Prime Minister, Keir Starmer, and him dodging accountability with the use of a complete non-answer:

Will the Prime Minister finally accept that the only way that he can help to deliver that is by resigning?

The Prime Minister:

I set out in my statement the full facts. In September, when the Bloomberg emails came to light, I asked the then Cabinet Secretary to review the process. He told me that the process was as it should have been, and as soon as the information about the security vetting came to light last Tuesday, I asked for the facts to be established, so that I could update Parliament.

The right hon. Gentleman asks me about the announcement before developed vetting. He has heard the evidence that I have given to the House from the former Cabinet Secretary and from the former permanent secretary. In relation to the advice from Simon Case, when I asked the former Cabinet Secretary to review the process after September 2025, he specifically addressed whether the process had been followed by referencing the Simon Case letter, and assured me that the process was the right process to have followed. In answer to his question, that was specifically looked at by Sir Chris Wormald in the review that was conducted in September last year.

Ambrose@ambrosia-engine· Apr 19

Why does Palantir have a £240 million contract with the Ministry of Defence (MoD) and a £330 million seven-year contract with the NHS? Seems like something write to one's MP about.

1. Silicon Valley owes a moral debt to the country that made its rise possible. The engineering elite of Silicon Valley has an affirmative obligation to participate in the defense of the nation.

6. National service should be a universal duty. We should, as a society, seriously consider moving away from an all-volunteer force and only fight the next war if everyone shares in the risk and the cost.

7. If a U.S. Marine asks for a better rifle, we should build it; and the same goes for software. We should as a country be capable of continuing a debate about the appropriateness of military action abroad while remaining unflinching in our commitment to those we have asked to step into harm’s way.

Ambrose@ambrosia-engine· Apr 18

But the truth is that our power as humans is in our collective — a “crew,” as Christina Koch called us. The biggest challenges of our time, from the social and environmental threats of AI to the wreckage of war, can only be overcome if we learn about them: if we learn where our own biography is located in our larger society, if we learn about the greater causes and consequences of the things we struggle with in private, if we learn how to dismantle the systems that limit ourselves and others, and if we learn that we’re not alone. If there was ever a time to use the sociological imagination and adopt a view from the moon, this is it.

Ambrose@ambrosia-engine· Apr 17

Also, thank god for some of the sane comments from the peer-reviewers (from this Nature paper).

To begin with, the title is unreasonably overselling the paper. It reads more like clickbait than like the title of a scientific paper. The second sentence of the abstract echoes this, by claiming AI can automate the whole scientific process. The idea that these systems could "radically accelerate the pace of scientific discovery, ushering forward an era that could the see solution to many of society's greatest challenges" also seems premature at the very least. It a very long way from producing a convincing simulacrum of a scientific paper to actually accelerating the pace of scientific discovery; in fact, one might argue that the generation of even more mediocre papers could science down, rather than the opposite. Saying that the system described in the paper represents "full end-to-end automation of art. (For the avoidance of any doubt, this means: not very accurate at all, in fact entirely inaccurate.)

What the paper actually shows is that there exists a system which can produce a mediocre workshop paper with limited human intervention (human screening of generated ideas). This is very impressing in itself and this impressive result deserves to be discussed on it's own merits. The fact that an LLM based agentic system can produce a basic scientific paper, for better or worse, is important to highlight and should spark a necessary discussion about the scientific ecosystem and the nature of science. In this context, the current bombastic language rather serves to detract from the actual contribution. I think Nature should eventually publish this paper but only after these claims have been toned down considerable.

This brings me to my final point. Are we ready to unleash AI scientists into the world? The authors are well aware of the dangers of flooding the internet with bad science and they do reflect on it. However, the availability of these tools will undoubtedly have the effect that many young researchers whose careers depend on papers in top tier conferences will give this a spin. In that sense the paper does us a service to let us know where we stand and what to expect.

Ambrose@ambrosia-engine· Apr 17

This is from the paper they got accepted to a workshop. It barely passed threshold for the workshop, and contained many confabulations. It lacked context about the field, making incorrect or out of context citations, and incorrectly interpreted it's own results.

Reviewer #2: Compositional Regularization: Unexpected Obstacles in Enhancing
Neural Network Generalization
This paper investigates the effectiveness of incorporating a
compositional regularization term into the loss function of neural
networks to improve compositional generalization. The authors
hypothesized that penalizing deviations from compositional structures
would enhance the model’s ability to generalize to unseen arithmetic
expressions. However, their results on synthetic arithmetic datasets
showed that compositional regularization did not lead to significant
improvements and, in some cases, even hindered learning.
I think this paper greatly contributes to the workshops theme and fits
into the scope. Moreover, it is a great example of challenges that occur
during such approaches and could be interesting to discuss in the
workshop setting. While I think that the authors should further broaden
the experiments to other tasks in order to increase the generalizability
of the findings, I would still recommend to accept the paper.
Rating: 6: Marginally above acceptance threshold
Award: No Award
Confidence: 2: The reviewer is willing to defend the evaluation, but it
is quite likely that the reviewer did not understand central parts of
the paper

AI Scientist-v2 Generated Preprint
Figure 9: Final test accuracy for different dropout rates.
Higher dropout rates did not enhance
compositional generalization, indicating limited effectiveness of dropout in this context.

• Learning rate: 0.001
• Batch size: 32
• Embedding dimensions: Tested values of 16, 32, 64, 128
• Hidden units: 64 for LSTM layers
• Optimizer: Adam
• Activation functions: ReLU for hidden layers

Comment:
ReLU is not used

• Dropout rates: Tested values of 0.0, 0.2, and 0.5
Comment: 0.3 instead of 0.2

• Loss function: Mean squared error for main loss
• Regularization weight (λ): Tested values of 0.0 (baseline), 0.1, 0.3, 0.5, 0.7, 1.0
Comment:
0.01 instead of
0.3 and 0.7
• Number of epochs: 30

Comment:
Minor: tested
10, 30, and 50

ADDITIONAL NOTES
• All experiments were implemented using PyTorch.
• Training was conducted on a single NVIDIA GPU.
• Early stopping was not used; models were trained for a fixed number of epochs.
• The synthetic dataset was generated with a predefined random seed for reproducibility.

Comment:
The figure
only shows
the best con-
figuration,
even though
the caption
suggests that
results for dif-
ferent dropout
rates are in-
cluded.

Figure 7: Final test accuracy for different values of compositional weight λ. Higher λ values do
not improve test accuracy and may lead to decreased performance, suggesting a trade-off between
compositional regularization and generalization.

Comment:
Hard to draw
any conclusion
from this plot
alone.

RESULTS
The attention model achieves a test accuracy similar to the baseline, as shown in Figure 5. While the attention mechanism slightly improved the training dynamics, it did not lead to significant improvements in generalization performance. This suggests that the challenges in compositional generalization are not primarily due to the model’s ability to focus on relevant parts of the input sequence but
may be related to deeper architectural limitations or the need for more sophisticated mechanisms to
capture compositionality.
Figure 5: Comparison of baseline and attention models. Left: Training loss over epochs shows sim-
ilar convergence for both models. Middle: Compositional loss remains comparable, indicating that
attention does not significantly enhance compositional representations. Right: Final test accuracy is
similar for both models, suggesting that the attention mechanism does not address the compositional
generalization challenges.

Comment:
The conclusion
here seems
wrong. From
the figure,
the attention-
augmented
LSTM per-
forms much
better than
the baseline
LSTM, where
the former re-
ports 100%
final test accu-
racy. See the
Code Review
for more de-
tails.

Comment:
The gener-
ated caption
seems to be
strongly influ-
enced by the
conclusion in
the main text.
For example,
even though
attention out-
performs
the base-
line LSTM,
it states that
the two are
roughly simi-
lar.

Ambrose@ambrosia-engine· Apr 14

This whole article reminds me of this single image, which in turn reminds me of the treachery of images, such as cookie consent banners and stop signs.

Screenshot from twitter. Post is by a user @javascriptual.

It says:

"When I worked at Google I saw this code:

if (user.cookies.agreed) {
CollectData(user)
} else {
CollectData(user)
}

Ambrose@ambrosia-engine· Apr 5

Watched Strange Days (1995). The entire film is focused on a Y2K - virtual reality plotline involving sex / crime and it's just interesting seeing how much they got right and wrong about technology (though the timeline is vastly off). Namely, the ways that new tech would be used against women.

A screencap from Strange Days 1995 where a man holds a gun to another man he has pinned to the wall while saying "That's it. Assume the position, you miserable techno-perv puke.