Zotero Catalogue

Inductive learning—that is, learning a new concept or category by observing exemplars—happens constantly, for example, when a baby learns a new word or a doctor classifies x-rays. What influence does the spacing of exemplars have on induction? Compared with massing, spacing enhances long-term recall, but we expected spacing to hamper induction by making the commonalities that define a concept or category less apparent. We asked participants to study multiple paintings by different artists, with a given artist's paintings presented consecutively (massed) or interleaved with other artists' paintings (spaced). We then tested induction by asking participants to indicate which studied artist (Experiments 1a and 1b) or whether any studied artist (Experiment 2) painted each of a series of new paintings. Surprisingly, induction profited from spacing, even though massing apparently created a sense of fluent learning: Participants rated massing as more effective than spacing, even after their own test performance had demonstrated the opposite.

At least two temporally and mechanistically distinct forms of memory are conserved across many species: short-term memory that persists minutes to hours after training and long-term memory (LTM) that persists days or longer. In general, repeated training trials presented with intervening rest intervals (spaced training) is more effective than massed training (the same number of training trials presented with no or short intervening rest intervals) in producing LTM. LTM requires de novo protein synthesis, and cAMP response element-binding protein (CREB) may be one of the transcription factors regulating the synthesis of new proteins necessary for the formation of LTM. Here we show that rats given massed fear conditioning training show no or weak LTM, as measured by fear-potentiated startle, compared with rats given the same amount of training but presented in a spaced manner. Increasing CREB levels specifically in the basolateral amygdala via viral vector-mediated gene transfer significantly increases LTM after massed fear training. The enhancing effect of CREB overexpression on LTM formation is shown to be specific in terms of biochemistry, anatomy, time course, and the training procedure used. These results suggest that CREB activity in the amygdala serves as a molecular switch for the formation of LTM in fear conditioning.

In statistics, an effect size is a quantitative measure of the magnitude of a phenomenon. It can refer to the value of a statistic calculated from a sample of data, the value of one parameter for a hypothetical population, or the equation that operationalizes how statistics or parameters lead to the effect size value. Examples of effect sizes include the correlation between two variables, the regression coefficient in a regression, the mean difference, and the risk of a particular event (such as a heart attack). Effect sizes are a complementary tool for statistical hypothesis testing, and play an important role in statistical power analyses to assess the sample size required for new experiments. Effect size calculations are fundamental to meta-analysis, which aims to provide the combined effect size based on data from multiple studies. The group of data-analysis methods concerning effect sizes is referred to as estimation statistics. Effect size is an essential component in the evaluation of the strength of a statistical claim, and it is the first item (magnitude) in the MAGIC criteria. The standard deviation of the effect size is of critical importance, as it indicates how much uncertainty is included in the observed measurement. A standard deviation that is too large will make the measurement nearly meaningless. In meta-analysis, which aims to summarize multiple effect sizes into a single estimate, the uncertainty in studies' effect sizes is used to weight the contribution of each study, so larger studies are considered more important than smaller ones. The uncertainty in the effect size is calculated differently for each type of effect size, but generally only requires knowing the study's sample size (N), or the number of observations (n) in each group. Reporting effect sizes or estimates thereof (effect estimate [EE], estimate of effect) is considered good practice when presenting empirical research findings in many fields. The reporting of effect sizes facilitates the interpretation of the importance of a research result, in contrast to its statistical significance. Effect sizes are particularly significant in social science and medical research, with the latter emphasizing the importance of the magnitude of the average treatment effect. Effect sizes may be measured in relative or absolute terms. In relative effect sizes, two groups are directly compared with each other, as in odds ratios and relative risks. A larger absolute value always indicates a stronger effect for absolute effect sizes. Many types of measurements can be expressed as either absolute or relative, and these can be used together because they convey different information. A prominent task force in the psychology research community made the following recommendation: Always present effect sizes for primary outcomes...If the units of measurement are meaningful on a practical level (e.g., number of cigarettes smoked per day), then we usually prefer an unstandardized measure (regression coefficient or mean difference) to a standardized measure (r or d).

The Graduate Record Examinations (GRE) is a standardized test that is part of the admissions process for some graduate schools in the United States and Canada. The GRE is owned and administered by the Educational Testing Service (ETS). The test was established in 1936 by the Carnegie Foundation for the Advancement of Teaching. According to ETS, the GRE General Test aims to measure verbal reasoning, quantitative reasoning, analytical writing, and critical thinking skills that have been acquired over a long period of learning. The content of the GRE consists of certain specific data analysis or interpretation, arguments and reasoning, algebra, geometry, arithmetic, and vocabulary sections. The GRE General Test is offered as a computer-based exam administered at testing centers and institution owned or authorized by Prometric as well as at home. In the graduate school admissions process, the level of emphasis that is placed upon GRE scores varies widely among schools and departments. The importance of a GRE score can range from being a mere admission formality to an important selection factor. The GRE was significantly overhauled in August 2011, resulting in an exam that is adaptive on a section-by-section basis, rather than question by question, so that the performance on the first verbal and math sections determines the difficulty of the second sections presented (excluding the experimental section). Overall, the test retained the sections and many of the question types from its predecessor, but the scoring scale was changed to a 130 to 170 scale (from a 200 to 800 scale). The cost to take the test is US$220, although ETS will reduce the fee under certain circumstances. It also provides financial aid to GRE applicants who prove economic hardship. ETS does not release scores that are older than five years, although graduate program policies on the acceptance of scores older than five years will vary. Once almost universally required for admission to Ph.D. science programs in the U.S., its use for that purpose has fallen precipitously since 2020.

In economics, hyperbolic discounting is a time-inconsistent model of delay discounting. It is one of the cornerstones of behavioral economics and its brain-basis is actively being studied by neuroeconomics researchers. According to the discounted utility approach, intertemporal choices are no different from other choices, except that some consequences are delayed and hence must be anticipated and discounted (i.e., reweighted to take into account the delay). Given two similar rewards, humans show a preference for one that arrives in a more prompt timeframe. Humans are said to discount the value of the later reward, by a factor that increases with the length of the delay. In the financial world, this process is normally modeled in the form of exponential discounting, a time-consistent model of discounting. Many psychological studies have since demonstrated deviations in instinctive preference from the constant discount rate assumed in exponential discounting. Hyperbolic discounting is an alternative mathematical model that agrees more closely with these findings. According to hyperbolic discounting, valuations fall relatively rapidly for earlier delay periods (as in, from now to one week), but then fall more slowly for longer delay periods (for instance, more than a few days). For example, in an early study subjects said they would be indifferent between receiving $15 immediately or $30 after 3 months, $60 after 1 year, or $100 after 3 years. These indifferences reflect annual discount rates that declined from 277% to 139% to 63% as delays got longer. This contrasts with exponential discounting, in which valuation falls by a constant factor per unit delay and the discount rate stays the same. The standard experiment used to reveal a test subject's hyperbolic discounting curve is to compare short-term preferences with long-term preferences. For instance: "Would you prefer a dollar today or three dollars tomorrow?" or "Would you prefer a dollar in one year or three dollars in one year and one day?" It has been claimed that a significant fraction of subjects will take the lesser amount today, but will gladly wait one extra day in a year in order to receive the higher amount instead. Individuals with such preferences are described as "present-biased". The most important consequence of hyperbolic discounting is that it creates temporary preferences for small rewards that occur sooner over larger, later ones. Individuals using hyperbolic discounting reveal a strong tendency to make choices that are inconsistent over time – they make choices today that their future self would prefer not to have made, despite knowing the same information. This dynamic inconsistency happens because hyperbolas distort the relative value of options with a fixed difference in delays in proportion to how far the choice-maker is from those options.

Edward Oakley Thorp (born August 14, 1932) is an American mathematics professor, author, hedge fund manager, and blackjack researcher. He pioneered the modern applications of probability theory, including the harnessing of very small correlations for reliable financial gain. Thorp is the author of Beat the Dealer, which mathematically proved that the house advantage in blackjack could be overcome by card counting. He also developed and applied effective hedge fund techniques in the financial markets, and collaborated with Claude Shannon in creating the first wearable computer. Thorp received his Ph.D. in mathematics from the University of California, Los Angeles in 1958, and worked at the Massachusetts Institute of Technology (MIT) from 1959 to 1961. He was a professor of mathematics from 1961 to 1965 at New Mexico State University. He then joined the University of California, Irvine where he was a professor of mathematics from 1965 to 1977 and a professor of mathematics and finance from 1977 to 1982.

Neuroprosthetics (also called neural prosthetics) is a discipline related to neuroscience and biomedical engineering concerned with developing neural prostheses. They are sometimes contrasted with a brain–computer interface, which connects the brain to a computer rather than a device meant to replace missing biological functionality. Neural prostheses are a series of devices that can substitute a motor, sensory or cognitive modality that might have been damaged as a result of an injury or a disease. Cochlear implants provide an example of such devices. These devices substitute the functions performed by the eardrum and stapes while simulating the frequency analysis performed in the cochlea. A microphone on an external unit gathers the sound and processes it; the processed signal is then transferred to an implanted unit that stimulates the auditory nerve through a microelectrode array. Through the replacement or augmentation of damaged senses, these devices are intended to improve the quality of life for those with disabilities. These implantable devices are also commonly used in animal experimentation as a tool to aid neuroscientists in developing a greater understanding of the brain and its functioning. By wirelessly monitoring the brain's electrical signals sent out by electrodes implanted in the subject's brain, the subject can be studied without the device affecting the results. Accurately probing and recording the electrical signals in the brain would help better understand the relationship among a local population of neurons that are responsible for a specific function. Neural implants are designed to be as small as possible in order to be minimally invasive, particularly in areas surrounding the brain, eyes, or cochlea. These implants typically communicate with their prosthetic counterparts wirelessly. Additionally, power is currently received through wireless power transmission through the skin. The tissue surrounding the implant is usually highly sensitive to temperature rise, meaning that power consumption must be minimal in order to prevent tissue damage. The neuroprosthetic currently undergoing the most widespread use is the cochlear implant, with over 736,900 in use worldwide as of 2019.

Efficient memorization using the spacing effect: literature review of widespread applicability, tips on use & what it’s good for.

A description of advanced tips and tricks for effective Internet research of papers/books, with real-world examples.

Fisher's fundamental theorem of natural selection is an idea about genetic variance in population genetics developed by the statistician and evolutionary biologist Ronald Fisher. The proper way of applying the abstract mathematics of the theorem to actual biology has been a matter of some debate, however, it is a true theorem. It states: "The rate of increase in fitness of any organism at any time is equal to its genetic variance in fitness at that time." Or in more modern terminology: "The rate of increase in the mean fitness of any organism, at any time, that is ascribable to natural selection acting through changes in gene frequencies, is exactly equal to its genetic variance in fitness at that time".

A physics-first view of the principles underlying effective accelerationism

Verbatim repost of original. Authored by @BasedBeffJezos and @bayeslord

“AI has not created new risks. It has made old risks urgent.”

Sweden goes to the polls on today in an extremely close general election that could see the far right enter government for the first time if the country’s rightwing parties can form a majority.&nbs…

This Gradient Updates issue investigates the economic consequences of fully automating remote work.

The US economy continues to expand faster than other G7 economies, but the driver is the humungous investment in AI models, data centres and all the AI-related chips and technology.  The US co…

NOTE: Before you read these instructions, think about whether dubbing or recreating is the best choice for your language and local audience. Khan Academy allows its translation teams to decide wh...

(a tiresome thing from the internet this weekend, and some new stories)

Locally Optimal Strategies (v3, 2026 Sep 23)

I recently had an experience while doing a simple Google search that was so profoundly weird that it stopped me in my tracks. Understanding this Google se...

Heritability estimates of general intelligence in adulthood generally range from 75 to 85%, with all heritability due to additive genetic influences, while genetic dominance and shared environmental factors are absent, or too small to be detected. These estimates are derived from studies based on the classical twin design and are based on the assumption of random mating. Yet, considerable positive assortative mating has been reported for general intelligence. Unmodeled assortative mating may lead to biased estimates of the relative magnitude of genetic and environmental factors. To investigate the effects of assortative mating on the estimates of the variance components of intelligence, we employed an extended twin-family design. Psychometric IQ data were available for adult monozygotic and dizygotic twins, their siblings, the partners of the twins and siblings, and either the parents or the adult offspring of the twins and siblings (N = 1314). Two underlying processes of assortment were considered: phenotypic assortment and social homogamy. The phenotypic assortment model was slightly preferred over the social homogamy model, suggesting that assortment for intelligence is mostly due to a selection of mates on similarity in intelligence. Under the preferred phenotypic assortment model, the variance of intelligence in adulthood was not only due to non-shared environmental (18%) and additive genetic factors (44%) but also to non-additive genetic factors (27%) and phenotypic assortment (11%).This non-additive nature of genetic influences on intelligence needs to be accommodated in future GWAS studies for intelligence.

The first time AI makes me tear up I'm biking down a sunny Brooklyn street.

Bibliography for directory 𝚜𝚝𝚊𝚝𝚒𝚜𝚝𝚒𝚌𝚜/𝚜𝚝𝚢𝚕𝚘𝚖𝚎𝚝𝚛𝚢/𝚝𝚛𝚞𝚎𝚜𝚒𝚐𝚑𝚝, most recent first: 3 related tags, 45 annotations, & 40 links (parent).

A guest post by Lars Doucet

The Royal Spanish Football Federation (RFEF) has announced an agreement for its players to freeze their eggs and undergo fertility treatment.

The changing face of Quebec is showing up in an unexpected place: the diaper aisle. Major pharmacy chains in Quebec are now selling more adult incontinence products than diapers for children, according to information obtained by Radio-Canada. A shift on pharmacy shelves The comparison was documented using sales information from pharmacy chains Jean Coutu and […]

How my high risk pregnancy has influenced my thinking about fertility declines

What do Pope Leo XIV and Billie Eilish have in common? Neither have kids, and both were spontaneously named by numerous people in a recent survey as the person they most admired in the world. Among that set of popular celebrities, their admirers had some of the lowest desired family sizes: 2.04 kids for Pope Leo’s fans, 2.1 for Eilish’s. At the opposite end of the spectrum, superfans of Cardi B (4 kids), Tom Brady (3 kids), Michael Jordan (5 kids), and Chris Brown (3 kids) topped the list for desired family size, with all their fans wanting 3.2 or more children.The small family sizes for the pope’s fans may be a misleading number; some of the people who named him as their favorite public figure might be Catholic clergy, committed to celibacy. When we add a range of demographic controls to adjust for this effect, the Pope Leo Fanclub’s adjusted desired family size rises to 2.18, while Eilish’s falls to 1.72, the lowest of any celebrity we studied.Those sound like two cherry-picked examples. They aren’t. For the Institute for Family Studies’ 2026 State of Fertility report, my colleagues and I surveyed more than 4,700 Americans about the texture of their social and cultural lives, and one finding jumped off the page. We asked every respondent to name the public figure they most admired, then matched that figure to the number of children they were publicly known to have, as well as other demographic traits. The result: Each additional child borne by someone’s most-admired celebrity was associated with larger family-size desires for the fan — about as strong an association as the number of siblings that person grew up with.For women, the celebrity effect was larger and more statistically reliable. The effect of being a fan of a big-family celebrity on desired family size was nearly as large as the effect of being a regular churchgoing person, and I think we can safely assume the median Cardi B superfan is not the median churchgoer.These aren’t small effects, and they appear in a country where the birth rate is below 1.6 children and falling. It’s high time we got serious about asking what exactly it is about modern culture that leads to such low birth rates. It’s not just simple economics: Religious people have far more kids than non-religious yet work in the same economy. While pro-natal policies that give cash to parents do boost births a bit, the overall effect has been a bit underwhelming. It’s time to take a good look at our culture, and culture, in a country like ours, is set at the top. It is set by the famous across various dimensions of society.There are exceptions to the celebrity effect, of course. NBA YoungBoy and Elon Musk both have around dozen acknowledged children, yet their fans have below-average desired family sizes: around 2.2-2.3 children in both cases (and Elon’s under-30 fans only desire about one child on average). In both cases, those are men who had children with many women, which might undermine the influence of their big families on their fans. Most people don’t just want offspring; they want a family — a spouse who loves them, children around them, and a home full of memories. Thus, it’s not just celebrities who have many children who may shape fertility desires, but celebrities who have compelling family lives for fans to emulate.This shouldn’t surprise anyone who has watched a teenage girl reorganize her wardrobe around a pop star or a young man pick up lifting because an athlete made it look like the point of being alive. We are imitative creatures. We always have been. We rehearse adulthood by copying the adults we find magnetic. Nowadays, the adults many people find most magnetic — the ones whose lives fit into our pockets for hours a day — are, disproportionately, people whose brand is built on everything except raising children.This is not a knock on any individual star. It’s a description of an incentive structure. In some cases, it’s explicit: K-Pop stars are often required by their contracts to remain single and childless, and if by chance they do start a family, their fans often turn on them. They exist to be accessible heartthrobs, not real people.Fame in 2026 rewards availability, mobility, reinvention, and the appearance of perpetual youth. It requires a carefully cultivated and sustained personal brand meant to capture a devoted audience, not just the highest number of eyeballs.Unlike the celebrities we see on TV, parents have to set limits on their availability, and they often put down roots. And, well, let’s face it: It’s called “dad bod” for a reason. Sure, a celebrity with one carefully managed child and a glam squad might read as aspirational, but a celebrity with five? Their kids are almost certainly not on screen, unless the celebrity deliberately puts them in the frame. In turn, the culture-shaping class in our society tends either to be childless or to present themselves as such to maintain their image. As we find in the data, the result is that their fans suppress their own fertility desires.Not all hope is lost. The same cultural machinery that has talked us out of children can talk us back into them. We don’t need to lecture young people about their duty to the nation. Nobody has ever been nagged into wanting a baby. We need to make family life visible, attractive, and high-status again, and the fastest way to do that is through the people we already can’t stop watching. We need more famous families.So how do you manufacture a more familistic celebrity culture? You can’t draft people into parenthood, and you shouldn’t try. But you can stop treating elite fertility as a private eccentricity and start treating it as a public good worth incentivizing. France already does a version of this: Its family quotient system hands its largest tax breaks to high-earning parents — on the theory that the people who set the tone for society should find children easy and rewarding rather than punishing. We could do the same: make it absurdly, conspicuously rewarding for the most influential Americans to have big families and let the cultural spillover do its work on the millions watching.It is possible that paying Taylor Swift a billion dollars per kid (and maybe a billion more for a nice lullaby album) would do more for the American birth rate than spending that same billion on another round of child tax credits spread out, with a few bucks extra for each kid in America. I am not actually proposing we cut Taylor a check, but some math from our estimated effect sizes suggests that Taylor Swift having two kids could theoretically boost fertility for her cohort of fans by over 200,000 babies. If we take the effect sizes from published studies of cash benefits as gospel, a billion dollars spent on baby bonuses for regular people would only get you about 40,000 extra babies. Subsidizing celebrities’ families might be a bad idea for a whole host of reasons (and probably no image-conscious celebrity would ever agree to take the money!), but the principle of the matter stands: Cultural interventions could be valuable ways to reverse the birth dearth.Of course, family aspirations aren’t the same as family outcomes. We find that celebrity families have a big impact on the total number of children young people want to have; their effect on couples’ actual intentions to have more is smaller. It takes more than a glossy Instagram family to change what people actually do. Turning the desire for kids into the third kid in the car seat requires something else: friends who show up, neighbors who help, and a society that makes room for parents.In the same study showing celebrity fertility effects, we found that individuals who had more supportive friends (i.e., whose friends would deliver them meals after a baby, celebrate with them, or babysit a child) not only wanted more kids — like the celebrity effect — but also actually intended to have more kids. Celebrities can help ripen the culture for family life, but supportive American communities remain indispensable to making families’ desires a reality.***Lyman Stone is a senior fellow and director of the Pronatalism Initiative at the Institute for Family Studies. He holds a Ph.D. in sociology from McGill University and is also director of research at the consulting firm Demographic Intelligence.This article is part of Upstream, The Daily Wire’s new home for culture and lifestyle. Real human insight and human stories — from our featured writers to you.

Scientific progress is a key driver of economic growth and prosperity. There is great excitement - but also concerns - about the impacts of AI on science, but so far little data. We provide early insights on this from three data sources: a sample of 15 million Gemini interactions, an inventory of over 2,600 specialized AI models across disciplines, and a survey of over 600 scientists. We map these data to a new taxonomy of scientific tasks to study how scientists are using AI. Four main findings emerge. First, we find broad adoption and coverage: scientists use AI more than most other occupations. Specialized AI models have broad disciplinary coverage and are highly cited. Nearly half of the scientists surveyed report using some form of AI every day. Second, we document evidence that LLMs (proxied through Gemini usage) and specialized models act as complements—LLMs are used for general analysis, coding, and manuscript preparation, while specialized models provide domain-specific predictions, data generation and classification. Third, scientists report large productivity gains from using AI: a saving of nearly 7 hours per week, time which is primarily re-invested in more research. Finally, we show that AI is already changing the scientific process. As some stages of scientific research become easier, bottlenecks shift downstream. Scientists report an increased backlog of untested hypotheses and substantial demand for output verification. Our findings suggest that AI holds significant potential to increase scientific productivity. However, as with other sectors, its ultimate impact will be governed by complex task interdependencies and investment into the elimination of emerging bottlenecks.

Studies estimating the cumulative risk of imprisonment have been central to documenting the rise of incarceration and its consequences for inequality in the United States. Data limitations have stymied efforts to document the risk of criminal legal contact falling short of imprisonment, however, and few studies estimate risks among groups other than Black and White men. Although less severe than imprisonment, subfelony contact also carries damaging and lifelong consequences. Using more than 25 years of administrative criminal record data from Texas, this study estimates the prevalence of both subfelony and felony arrest, prosecution, conviction, and incarceration across five cohorts of Black, White, and Hispanic men and women born between 1971 and 1995. Findings show that subfelony contact is 3‒18 times as prevalent, and far more widely distributed across the population, as felony conviction and incarceration: fewer than 3% of Black women go to prison by age 24, but at least 22% are arrested at least once. Although risks of felony events have fallen substantially for some groups, subfelony criminalization has persisted at high and increasingly unequal rates. These findings have implications for our understanding of the reach of the criminal legal system and its role in shaping inequality in the twenty-first century.

An exploration of a theory of marital decline

Large-scale agentic training and evaluation with large language models (LLMs) rely on isolated, stateful execution environments in which models inspect repositories, invoke tools, execute commands, and interact with task-specific services. These workloads create sandboxes in large bursts, span heterogeneous functionality and isolation requirements, retain state across long interactions, and draw from large image corpora with limited reuse. Supporting them therefore requires an elastic execution platform rather than a single sandbox runtime. This report presents DeepSeek Elastic Compute (DSec), a production sandbox platform that exposes FnCall, container, microVM, and full-VM sandbox backends through a unified SDK. DSec coordinates placement and lifecycle management across the cluster, composes environments from independently versioned layers, combines memory sharing, reclamation, and CPU scheduling for high-density execution, and loads image data on demand from Fire-Flyer File System (3FS), a cluster-wide distributed filesystem. DSec is co-designed with the reinforcement learning (RL) framework, decouples stateful rollout execution from preemptible GPU training, coordinates sandbox lifecycle with training to preserve rollout state while reclaiming idle resources, and mitigates agent misbehavior such as reward hacking. A single production-scale unit of DSec spans around 160 nodes, serving about 3 million sandboxes per day; in production, it supports over 380,000 concurrent sandboxes and sustains over 5,000 sandbox creations per second. Our evaluation and deployment experience show that these mechanisms reduce environment setup and image-distribution overhead, improve memory efficiency, and preserve latency-sensitive performance under high-density overcommit.

[This is a guest post by Amit Sahai. This blog post was initially written in a different file format and converted using AI. — T.] When I was an undergraduate student, I remember talking with…

The Digital Immune System: Rethinking Quantum Error Correction

We are past the event horizon; the takeoff has started. Humanity is close to building digital superintelligence, and at least so far it’s much less weird than it seems like it should be. Robots...

Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition. The latest model developed by OpenAI, GPT-4, was trained using an unprecedented scale of compute and data. In this paper, we report on our investigation of an early version of GPT-4, when it was still in active development by OpenAI. We contend that (this early version of) GPT-4 is part of a new cohort of LLMs (along with ChatGPT and Google's PaLM for example) that exhibit more general intelligence than previous AI models. We discuss the rising capabilities and implications of these models. We demonstrate that, beyond its mastery of language, GPT-4 can solve novel and difficult tasks that span mathematics, coding, vision, medicine, law, psychology and more, without needing any special prompting. Moreover, in all of these tasks, GPT-4's performance is strikingly close to human-level performance, and often vastly surpasses prior models such as ChatGPT. Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system. In our exploration of GPT-4, we put special emphasis on discovering its limitations, and we discuss the challenges ahead for advancing towards deeper and more comprehensive versions of AGI, including the possible need for pursuing a new paradigm that moves beyond next-word prediction. We conclude with reflections on societal influences of the recent technological leap and future research directions.

Professor Chris Bishop is a Technical Fellow and Director at Microsoft Research AI4Science, in Cambridge. He is also Honorary Professor of Computer Science at the University of Edinburgh, and a Fellow of Darwin College, Cambridge. In 2004, he was elected Fellow of the Royal Academy of Engineering, in 2007 he was elected Fellow of the Royal Society of Edinburgh, and in 2017 he was elected Fellow of the Royal Society. Chris was a founding member of the UK AI Council, and in 2019 he was appointed to the Prime Minister’s Council for Science and Technology. At Microsoft Research, Chris oversees a global portfolio of industrial research and development, with a strong focus on machine learning and the natural sciences. Chris obtained a BA in Physics from Oxford, and a PhD in Theoretical Physics from the University of Edinburgh, with a thesis on quantum field theory. Chris's contributions to the field of machine learning have been truly remarkable. He has authored (what is arguably) the original textbook in the field - 'Pattern Recognition and Machine Learning' (PRML) which has served as an essential reference for countless students and researchers around the world, and that was his second textbook after his highly acclaimed first textbook Neural Networks for Pattern Recognition. Recently, Chris has co-authored a new book with his son, Hugh, titled 'Deep Learning: Foundations and Concepts.' This book aims to provide a comprehensive understanding of the key ideas and techniques underpinning the rapidly evolving field of deep learning. It covers both the foundational concepts and the latest advances, making it an invaluable resource for newcomers and experienced practitioners alike. Buy Chris' textbook here: https://amzn.to/3vvLcCh More about Prof. Chris Bishop: https://en.wikipedia.org/wiki/Christo... https://www.microsoft.com/en-us/resea... Support MLST: Please support us on Patreon. We are entirely funded from Patreon donations right now. Patreon supports get private discord access, biweekly calls, early-access + exclusive content and lots more.   / mlst   Donate: https://www.paypal.com/donate/?hosted... If you would like to sponsor us, so we can tell your story - reach out on mlstreettalk at gmail TOC: 00:00:00 - Intro to Chris 00:06:54 - Changing Landscape of AI 00:08:16 - Symbolism 00:09:32 - PRML 00:11:02 - Bayesian Approach 00:14:49 - Are NNs One Model or Many, Special vs General 00:20:04 - Can Language Models Be Creative 00:22:35 - Sparks of AGI 00:25:52 - Creativity Gap in LLMs 00:35:40 - New Deep Learning Book 00:39:01 - Favourite Chapters 00:44:11 - Probability Theory 00:45:42 - AI4Science 00:48:31 - Inductive Priors 00:58:52 - Drug Discovery 01:05:19 - Foundational Bias Models 01:07:46 - How Fundamental Is Our Physics Knowledge? 01:12:05 - Transformers 01:12:59 - Why Does Deep Learning Work? 01:16:59 - Inscrutability of NNs 01:18:01 - Example of Simulator 01:21:09 - Control