Morning Singularity Digest - 2026-04-24

Context

Submission history From: Michael Bernstein [view email][v1] Fri, 15 Nov 2024 11:14:34 UTC (2,928 KB) [v2] Wed, 22 Apr 2026 03:48:01 UTC (5,565 KB) Current browse context: cs.AI References & Citations Loading...

What's new

We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.

Key details

• We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).
• Agents predicted personality traits and behaviors in experiments with similar accuracy, and reduced disparities in accuracy across racial and ideological groups relative to demographics-only baselines.

Results & evidence

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).

Limitations / unknowns

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

What's new

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

Key details

• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).
• Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.
• As a featured case, we present Shell-Induced Behavioral Override (SIBO) -- a controlled experiment showing that Shell instructions categorically override a model's default cooperative behavior.

Results & evidence

• arXiv:2604.20871v1 Announce Type: cross Abstract: We introduce M-CARE (Model Clinical Assessment and Reporting for Evaluation), a clinical case report framework for AI model behavioral disorders adapted from human medicine.
• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Make entire codebase the context for any coding agent.

What's new

Make entire codebase the context for any coding agent.

Key details

• Make entire codebase the context for any coding agent.

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

This is something I always wanted to do, as I love workflowy.com, but I want to own my data.

I had tried a few times before, but could not until now with the help of AI.

This is actually the second try of AI assistance.

What's new

THe first failed completely as I did not do anything and simply let the agent go free.

Key details

• THe first failed completely as I did not do anything and simply let the agent go free.
• On this second attempt I was much more involved in the details and code organization.

  AMA

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Show HN: Safer – Sleep better while AI agents have shell access

What's new

Show HN: Safer – Sleep better while AI agents have shell access

Key details

• Show HN: Safer – Sleep better while AI agents have shell access

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Submission history From: Michael Bernstein [view email][v1] Fri, 15 Nov 2024 11:14:34 UTC (2,928 KB) [v2] Wed, 22 Apr 2026 03:48:01 UTC (5,565 KB) Current browse context: cs.AI References & Citations Loading...

What's new

We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.

Key details

• We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).
• Agents predicted personality traits and behaviors in experiments with similar accuracy, and reduced disparities in accuracy across racial and ideological groups relative to demographics-only baselines.

Results & evidence

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).

Limitations / unknowns

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

This is something I always wanted to do, as I love workflowy.com, but I want to own my data.

I had tried a few times before, but could not until now with the help of AI.

This is actually the second try of AI assistance.

What's new

THe first failed completely as I did not do anything and simply let the agent go free.

Key details

• THe first failed completely as I did not do anything and simply let the agent go free.
• On this second attempt I was much more involved in the details and code organization.

  AMA

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

What's new

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

Key details

• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).
• Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.
• As a featured case, we present Shell-Induced Behavioral Override (SIBO) -- a controlled experiment showing that Shell instructions categorically override a model's default cooperative behavior.

Results & evidence

• arXiv:2604.20871v1 Announce Type: cross Abstract: We introduce M-CARE (Model Clinical Assessment and Reporting for Evaluation), a clinical case report framework for AI model behavioral disorders adapted from human medicine.
• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Submission history From: Michael Bernstein [view email][v1] Fri, 15 Nov 2024 11:14:34 UTC (2,928 KB) [v2] Wed, 22 Apr 2026 03:48:01 UTC (5,565 KB) Current browse context: cs.AI References & Citations Loading...

What's new

We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.

Key details

• We test whether large language models (LLMs) can support a more general-purpose approach by building person-specific simulations (i.e., "generative agents") grounded in self-report data.
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).
• Agents predicted personality traits and behaviors in experiments with similar accuracy, and reduced disparities in accuracy across racial and ideological groups relative to demographics-only baselines.

Results & evidence

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...
• Using data from a diverse national sample of 1,052 Americans, we build agents from (i) two-hour, semi-structured interviews (elicited using the American Voices Project interview schedule), (ii) structured surveys (the General Social Survey and Big Five pers...
• On held-out General Social Survey items, agent accuracy reached 83% (interview only), 82% (surveys only), and 86% (combined) of participants' two-week test-retest consistency, compared with agents prompted only with individuals' demographics (74%).

Limitations / unknowns

• arXiv:2411.10109v2 Announce Type: replace Abstract: Machine learning can predict human behavior well when substantial structured data and well-defined outcomes are available, but these models are typically limited to specific outcomes and cannot readily be...

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

What's new

Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.

Key details

• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).
• Cases are organized into five categories: RLHF Performance Artifacts, Shell-Core Override Pathology, Context & Memory Conditions, Core Identity & Plasticity, and Stress, Methodology, & Boundary Conditions.
• As a featured case, we present Shell-Induced Behavioral Override (SIBO) -- a controlled experiment showing that Shell instructions categorically override a model's default cooperative behavior.

Results & evidence

• arXiv:2604.20871v1 Announce Type: cross Abstract: We introduce M-CARE (Model Clinical Assessment and Reporting for Evaluation), a clinical case report framework for AI model behavioral disorders adapted from human medicine.
• M-CARE provides a 13-section report format, a 4-axis diagnostic assessment system, and a nosological classification of AI behavioral conditions.
• We present 20 cases from three source categories: field observations of deployed agents (8), controlled experiments across three platforms (8), and published sources (4).

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

arXiv:2604.21082v1 Announce Type: cross Abstract: Training vision-language models (VLMs) for medical report generation is often hindered by the scarcity of high-quality annotated data.

What's new

In experiments on ophthalmological report generation, we show that this simple method improves efficiency across multiple data scales, achieving similar report quality with up to ten times less training data.

Key details

• This work evaluates the use of a weighted loss function to improve data efficiency.
• Compared to standard cross-entropy loss, which treats all token prediction errors equally, the reweighted loss shifts the focus to semantically salient tokens with outsized clinical importance.
• In experiments on ophthalmological report generation, we show that this simple method improves efficiency across multiple data scales, achieving similar report quality with up to ten times less training data.
• Computer Science > Computation and Language [Submitted on 22 Apr 2026] Title:Weighting What Matters: Boosting Sample Efficiency in Medical Report Generation via Token Reweighting View PDF HTML (experimental)Abstract:Training vision-language models (VLMs) fo...

Results & evidence

• arXiv:2604.21082v1 Announce Type: cross Abstract: Training vision-language models (VLMs) for medical report generation is often hindered by the scarcity of high-quality annotated data.
• Computer Science > Computation and Language [Submitted on 22 Apr 2026] Title:Weighting What Matters: Boosting Sample Efficiency in Medical Report Generation via Token Reweighting View PDF HTML (experimental)Abstract:Training vision-language models (VLMs) fo...

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

arXiv:2604.17628v2 Announce Type: replace Abstract: Wales' political landscape has been marked by growing accusations of bias in Welsh media.

What's new

This paper takes the first computational step toward testing those claims by examining Nation.Cymru, a prominent Welsh political news outlet.

Key details

• This paper takes the first computational step toward testing those claims by examining Nation.Cymru, a prominent Welsh political news outlet.
• I use a two-stage natural language processing (NLP) pipeline: (1) a robustly optimized BERT approach (RoBERTa) bias detector for efficient bias discovery and (2) a large language model (LLM) for target-attributed sentiment classification of bias labels from...
• A primary analysis of 15,583 party mentions across 2022-2026 news articles finds that Reform UK attracts biased framing at twice the rate of Plaid Cymru and over three times as negative in mean sentiment (p<0.001).
• A secondary analysis across four parties across both news and opinion articles shows that Plaid Cymru is the outlier, receiving markedly more favourable framing than any other party.

Results & evidence

• arXiv:2604.17628v2 Announce Type: replace Abstract: Wales' political landscape has been marked by growing accusations of bias in Welsh media.
• I use a two-stage natural language processing (NLP) pipeline: (1) a robustly optimized BERT approach (RoBERTa) bias detector for efficient bias discovery and (2) a large language model (LLM) for target-attributed sentiment classification of bias labels from...
• A primary analysis of 15,583 party mentions across 2022-2026 news articles finds that Reform UK attracts biased framing at twice the rate of Plaid Cymru and over three times as negative in mean sentiment (p<0.001).

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Study Reveals 75% of Enterprises Report Double-Digit AI Failure Rates

What's new

Study Reveals 75% of Enterprises Report Double-Digit AI Failure Rates

Key details

• Study Reveals 75% of Enterprises Report Double-Digit AI Failure Rates

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

South Korea police arrest man for posting AI photo of runaway wolf South Korean police have arrested a man for sharing an AI-generated image that misled authorities who were searching for a wolf that had broken out of a zoo in Daejeon city.

What's new

South Korea police arrest man for posting AI photo of runaway wolf South Korean police have arrested a man for sharing an AI-generated image that misled authorities who were searching for a wolf that had broken out of a zoo in Daejeon city.

Key details

• The 40-year-old unnamed man is accused of disrupting the search by creating and distributing a fake photo purporting to show Neukgu, the wolf, trotting down a road intersection.
• The photo, circulated hours after Neukgu went missing on 8 April, prompted authorities to urgently relocate their search operation, sending them on a wild wolf chase.
• The hunt for two-year-old Neukgu gripped the nation before he was finally caught near an expressway last week, nine days after his escape.
• The AI-generated image of Neukgu had prompted Daejeon city government to issue an emergency text to residents, warning them of a wolf near the intersection.

Results & evidence

• The 40-year-old unnamed man is accused of disrupting the search by creating and distributing a fake photo purporting to show Neukgu, the wolf, trotting down a road intersection.
• The photo, circulated hours after Neukgu went missing on 8 April, prompted authorities to urgently relocate their search operation, sending them on a wild wolf chase.
• Authorities are investigating him for disrupting government work by deception, an offence that carries up to five years in prison or a maximum fine of 10 million Korean won ($6,700; £5,000) For more than a week, the search for Neukgu captured the attention...

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

A New Framework for Evaluating Voice Agents (EVA)

What's new

A New Framework for Evaluating Voice Agents (EVA)

Key details

• A New Framework for Evaluating Voice Agents (EVA)

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

DeepSeek-V4: a million-token context that agents can actually use

What's new

DeepSeek-V4: a million-token context that agents can actually use

Key details

• DeepSeek-V4: a million-token context that agents can actually use

Results & evidence

• No hard numbers surfaced in the source text; treat claims as directional until benchmarks appear.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.

Context

Introducing GPT-5.5, our smartest model yet—faster, more capable, and built for complex tasks like coding, research, and data analysis across tools.

What's new

Introducing GPT-5.5, our smartest model yet—faster, more capable, and built for complex tasks like coding, research, and data analysis across tools.

Key details

• Introducing GPT-5.5, our smartest model yet—faster, more capable, and built for complex tasks like coding, research, and data analysis across tools.

Results & evidence

• Introducing GPT-5.5, our smartest model yet—faster, more capable, and built for complex tasks like coding, research, and data analysis across tools.

Limitations / unknowns

• Generalization outside curated tasks is still unclear.

Next-step validation checks

• Reproduce one claim with a public baseline and fixed evaluation settings.
• Check robustness on out-of-distribution or long-context cases.
• Track whether independent teams report matching results.