With public money, we are actually actively funding companies and startups thinking they will both create jobs and sell solutions to every problem. It is implied that every solution should be a technological one, should be sellable and should be intuitiv

Whether this "personalised advertisement" really works better than traditional one is up to debate. For Tim Hang, author of "Subtime Attention Crisis" and for Cory Doctorow, author of "How to destroy surveillance capitalism", the real impact on sales is negligible but as marketers think it works, they invest massive money in it, making the whole technology a very lucrative bubble.

Do public intellectuals even exist anymore? Do publics exist? Everywhere you go, the forces of power seek to atomize us away into marketable metrics, Facebook groups, random and randomized individuals, software-recognizable faces. Who are we, but clicks, trends, and lifestyles. We are everything but a public.

how do you help people who, like us, just never seem to have the time to figure this stuff out becase they're, like, suuuuper busy and stuff? You do it by showing them … the minumum helpful reminder at exactly the right time This is what I've called the "Just In Time" theory of user behavior for years. Sure, FAQs and tutorials and help centers are great and all, but who has the time for that? We're all perpetual intermediates here, at best.

a simple reminder at the time of the temptation is usually all it takes for people to suddenly "remember" their honesty.

These little white lies are the path of least resistance.

The experiments Ariely conducts prove again and again that most people will consistently and reliably cheat "just a little", to the extent that they can still consider themselves honest people. The gating factor isn't laws, penalties, or ethics. Surprisingly, that stuff has virtually no effect on behavior. What does, though, is whether they can personally still feel like they are honest people.

If you look at engineering as a system, how would you understand that system’s feedback loops?

engineering costs normalized against a core business metric (for example, cost to serve your most important API, calculated simply by dividing API requests per month by engineering spend for that month)

This paper presents COLT5 (ConditionalLongT5)

Over the past few years, many “efficient Trans-former” approaches have been proposed that re-duce the cost of the attention mechanism over longinputs (Child et al., 2019; Ainslie et al., 2020; Belt-agy et al., 2020; Zaheer et al., 2020; Wang et al.,2020; Tay et al., 2021; Guo et al., 2022). However,especially for larger models, the feedforward andprojection layers actually make up the majority ofthe computational burden and can render process-ing long inputs intractable

Still, we can look for telltale signs. Another symptom of memorization is that GPT is highly sensitive to the phrasing of the question. Melanie Mitchell gives an example of an MBA test question where changing some details in a way that wouldn’t fool a person is enough to fool ChatGPT (running GPT-3.5). A more elaborate experiment along these lines would be valuable.

In fact, we can definitively show that it has memorized problems in its training set: when prompted with the title of a Codeforces problem, GPT-4 includes a link to the exact contest where the problem appears (and the round number is almost correct: it is off by one). Note that GPT-4 cannot access the Internet, so memorization is the only explanation.

To benchmark GPT-4’s coding ability, OpenAI evaluated it on problems from Codeforces, a website that hosts coding competitions. Surprisingly, Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10 recent problems in the easy category. The training data cutoff for GPT-4 is September 2021. This strongly suggests that the model is able to memorize solutions from its training set — or at least partly memorize them, enough that it can fill in what it can’t recall.

Figure 3. The average drop in log probability (perturbation discrep-ancy) after rephrasing a passage is consistently higher for model-generated passages than for human-written passages. Each plotshows the distribution of the perturbation discrepancy d (x, pθ , q)for human-written news articles and machine-generated arti-cles; of equal word length from models GPT-2 (1.5B), GPT-Neo-2.7B (Black et al., 2021), GPT-J (6B; Wang & Komatsuzaki (2021))and GPT-NeoX (20B; Black et al. (2022)). Human-written arti-cles are a sample of 500 XSum articles; machine-generated textis generated by prompting each model with the first 30 tokens ofeach XSum article, sampling from the raw conditional distribution.Discrepancies are estimated with 100 T5-3B samples.

if we apply small perturbations to a passagex ∼ pθ , producing ̃x, the quantity log pθ (x) − log pθ ( ̃x)should be relatively large on average for machine-generatedsamples compared to human-written text.

As in prior work, we study a ‘white box’ setting (Gehrmannet al., 2019) in which the detector may evaluate the log prob-ability of a sample log pθ (x). The white box setting doesnot assume access to the model architecture or parameters.While most public APIs for LLMs (such as GPT-3) enablescoring text, some exceptions exist

Empirically, we find predictive entropy to be positively cor-related with passage fake-ness more often that not; there-fore, this baseline uses high average entropy in the model’spredictive distribution as a signal that a passage is machine-generated.

We find that supervised detectors can provide similardetection performance to DetectGPT on in-distribution datalike English news, but perform significantly worse than zero-shot methods in the case of English scientific writing andfail altogether for German writing. T

ex-tending DetectGPT to use ensembles of models for scoring,rather than a single model, may improve detection in theblack box setting

hile in this work, we use off-the-shelfmask-filling models such as T5 and mT5 (for non-Englishlanguages), some domains may see reduced performanceif existing mask-filling models do not well represent thespace of meaningful rephrases, reducing the quality of thecurvature estimate.

For models be-hind APIs that do provide probabilities (such as GPT-3),evaluating probabilities nonetheless costs money.

We simulate human re-vision by replacing 5 word spans of the text with samplesfrom T5-3B until r% of the text has been replaced, andreport performance as r varies.

Figure 5. We simulate human edits to machine-generated text byreplacing varying fractions of model samples with T5-3B gener-ated text (masking out random five word spans until r% of text ismasked to simulate human edits to machine-generated text). Thefour top-performing methods all generally degrade in performancewith heavier revision, but DetectGPT is consistently most accurate.Experiment is conducted on the XSum dataset

DetectGPT’s performancein particular is mostly unaffected by the change in languagefrom English to Germa

ecause the GPT-3 API does not provideaccess to the complete conditional distribution for each to-ken, we cannot compare to the rank, log rank, and entropy-based prior methods

improving detection offake news articles generated by 20B parameterGPT-NeoX

his approach, which we call DetectGPT,does not require training a separate classifier, col-lecting a dataset of real or generated passages, orexplicitly watermarking generated text. It usesonly log probabilities computed by the model ofinterest and random perturbations of the passagefrom another generic pre-trained language model(e.g, T5)

See ericmitchell.ai/detectgptfor code, data, and other project information.

The potential size of this market is hard to grasp — somewhere between all software and all human endeavors

Other hardware options do exist, including Google Tensor Processing Units (TPUs); AMD Instinct GPUs; AWS Inferentia and Trainium chips; and AI accelerators from startups like Cerebras, Sambanova, and Graphcore. Intel, late to the game, is also entering the market with their high-end Habana chips and Ponte Vecchio GPUs. But so far, few of these new chips have taken significant market share. The two exceptions to watch are Google, whose TPUs have gained traction in the Stable Diffusion community and in some large GCP deals, and TSMC, who is believed to manufacture all of the chips listed here, including Nvidia GPUs (Intel uses a mix of its own fabs and TSMC to make its chips).

Commoditization. There’s a common belief that AI models will converge in performance over time. Talking to app developers, it’s clear that hasn’t happened yet, with strong leaders in both text and image models. Their advantages are based not on unique model architectures, but on high capital requirements, proprietary product interaction data, and scarce AI talent. Will this serve as a durable advantage?

In natural language models, OpenAI dominates with GPT-3/3.5 and ChatGPT. But relatively few killer apps built on OpenAI exist so far, and prices have already dropped once.

Vertical integration (“model + app”). Consuming AI models as a service allows app developers to iterate quickly with a small team and swap model providers as technology advances. On the flip side, some devs argue that the product is the model, and that training from scratch is the only way to create defensibility — i.e. by continually re-training on proprietary product data. But it comes at the cost of much higher capital requirements and a less nimble product team.

Many apps are also relatively undifferentiated, since they rely on similar underlying AI models and haven’t discovered obvious network effects, or data/workflows, that are hard for competitors to duplicate.

We’re also not going deep here on MLops or LLMops tooling, which is not yet highly standardized and will be addressed in a future post.

Over the last year, we’ve met with dozens of startup founders and operators in large companies who deal directly with generative AI. We’ve observed that infrastructure vendors are likely the biggest winners in this market so far, capturing the majority of dollars flowing through the stack. Application companies are growing topline revenues very quickly but often struggle with retention, product differentiation, and gross margins. And most model providers, though responsible for the very existence of this market, haven’t yet achieved large commercial scale.

Here I’ve summarized Christian Tietze’s process, which I’m presently adopting / adapting:

You need to take a step back and form a picture of the overall structure of the ideas. Concretely, you might do that by clustering your scraps into piles and observing the structure that emerges. Or you might sketch a mind map or a visual outline.

Happiness is pushed to some later date in the future while your present self battles with the misery of the current moment.

Positive fantasies allow you to indulge in the desired future mentally…You can taste the sensations of what it’s like to achieve your goal in the present — this depletes your energy to pursue your desired future.

My goal was simply to scale this ladder over time. I worked the list 5 people at a time, starting at the bottom. I engaged relentlessly with those accounts until they noticed me and began engaging back.

Don’t try to convince everyone that what you say, feel, think, or have done is better than everyone else.

My goal with my content is to make it so recognizable that you would know it was me even if it didn't have my name on it. The same style. The same thought process. The same character.

Now, this can all be defeated with enough effort. For example, if you used another AI to paraphrase GPT’s output—well okay, we’re not going to be able to detect that. On the other hand, if you just insert or delete a few words here and there, or rearrange the order of some sentences, the watermarking signal will still be there. Because it depends only on a sum over n-grams, it’s robust against those sorts of interventions.

Anyway, we actually have a working prototype of the watermarking scheme, built by OpenAI engineer Hendrik Kirchner. It seems to work pretty well—empirically, a few hundred tokens seem to be enough to get a reasonable signal that yes, this text came from GPT. In principle, you could even take a long text and isolate which parts probably came from GPT and which parts probably didn’t.

So then to watermark, instead of selecting the next token randomly, the idea will be to select it pseudorandomly, using a cryptographic pseudorandom function, whose key is known only to OpenAI.

Eventually GPT will say, “oh, I know what game we’re playing! it’s the ‘give false answers’ game!” And it will then continue playing that game and give you more false answers. What the new paper shows is that, in such cases, one can actually look at the inner layers of the neural net and find where it has an internal representation of what was the true answer, which then gets overridden once you get to the output layer.

(3) A third direction, and I would say maybe the most popular one in AI alignment research right now, is called interpretability. This is also a major direction in mainstream machine learning research, so there’s a big point of intersection there. The idea of interpretability is, why don’t we exploit the fact that we actually have complete access to the code of the AI—or if it’s a neural net, complete access to its parameters? So we can look inside of it. We can do the AI analogue of neuroscience. Except, unlike an fMRI machine, which gives you only an extremely crude snapshot of what a brain is doing, we can see exactly what every neuron in a neural net is doing at every point in time. If we don’t exploit that, then aren’t we trying to make AI safe with our hands tied behind our backs?

“AI alignment”

And famously, self-driving cars have taken a lot longer than many people expected a decade ago. This is partly because of regulatory barriers and public relations: even if a self-driving car actually crashes less than a human does, that’s still not good enough, because when it does crash the circumstances are too weird. So, the AI is actually held to a higher standard. But it’s also partly just that there was a long tail of really weird events. A deer crosses the road, or you have some crazy lighting conditions—such things are really hard to get right, and of course 99% isn’t good enough here.

Okay, but one thing that’s been found empirically is that you take commonsense questions that are flubbed by GPT-2, let’s say, and you try them on GPT-3, and very often now it gets them right. You take the things that the original GPT-3 flubbed, and you try them on the latest public model, which is sometimes called GPT-3.5 (incorporating an advance called InstructGPT), and again it often gets them right. So it’s extremely risky right now to pin your case against AI on these sorts of examples! Very plausibly, just one more order of magnitude of scale is all it’ll take to kick the ball in, and then you’ll have to move the goal again.

If my interpretation of the Retrieval quadrant is correct, it will become much more difficult to be an average, or even above average, writer. Only the best will flourish. Perhaps we will see a rise in neo-generalists.

For many intellectual tasks, the people with the least skill overestimate themselves the most, a pattern popularly known as the Dunning–Kruger effect (DKE). The dominant account of this effect depends on the idea that assessing the quality of one's performance (metacognition) requires the same mental resources as task performance itself (cognition). Unskilled people are said to suffer a dual burden: they lack the cognitive resources to perform well, and this deprives them of metacognitive insight into their failings. In this Registered Report, we applied recently developed methods for the measurement of metacognition to a matrix reasoning task, to test the dual-burden account. Metacognitive sensitivity (information exploited by metacognition) tracked performance closely, so less information was exploited by the metacognitive judgements of poor performers; but metacognitive efficiency (quality of metacognitive processing itself) was unrelated to performance. Metacognitive bias (overall tendency towards high or low confidence) was positively associated with performance, so poor performers were appropriately less confident—not more confident—than good performers. Crucially, these metacognitive factors did not cause the DKE pattern, which was driven overwhelmingly by performance scores. These results refute the dual-burden account and suggest that the classic DKE is a statistical regression artefact that tells us nothing much about metacognition.

AI training data is filled with racist stereotypes, pornography, and explicit images of rape, researchers Abeba Birhane, Vinay Uday Prabhu, and Emmanuel Kahembwe found after analyzing a data set similar to the one used to build Stable Diffusion.

Throughout the 80s and 90s, private equity firms and hedge funds gobbled up local news enterprises to extract their real estate. They didn’t give a shit about journalism; they just wanted prime real estate that they could develop. And news organizations had it in the form of buildings in the middle of town. So financiers squeezed the news orgs until there was no money to be squeezed and then they hung them out to dry.

Perceptions of failure don’t always lead to shared ideas of how to learn from these lessons.

every country is going to need to reconsider its policies on misinformation. It’s one thing for the occasional lie to slip through; it’s another for us all to swim in a veritable ocean of lies. In time, though it would not be a popular decision, we may have to begin to treat misinformation as we do libel, making it actionable if it is created with sufficient malice and sufficient volume.

anyone skilled in the art can now replicate their recipe.

We test this hypothesis by training a predicted compute-optimal model, Chinchilla, that uses the same compute budget as Gopher but with 70B parameters and4× more more data. Chinchilla uniformly and significantly outperforms Gopher (280B), GPT-3 (175B),Jurassic-1 (178B), and Megatron-Turing NLG (530B) on a large range of downstream evaluation tasks.This also means that Chinchilla uses substantially less compute for fine-tuning and inference, greatlyfacilitating downstream usage. As a highlight, Chinchilla reaches a state-of-the-art average accuracy of67.5% on the MMLU benchmark, greater than a 7% improvement over Gopher

Amazon is hated on the right as a bulwark of progressivism. For instance, to pick a random example, GOP icon Tucker Carlson recently characterized the firm’s behavior as ‘modern-day book burning.’ And you can find an endless number of right-wing critiques. Conservatives distrust Amazon.

Whether you want to call them mottos, memes, or manifestos, words can be the building blocks of how we think and transmit ideas. You can also gauge how well someone is grasping your concepts—or at least making an effort to—by the language they’re responding to you with as well.

Of course, the closest you can get is having the activity available in your own living space, but as unused home treadmills and exercise bikes demonstrate, this has its pitfalls. There could be something about a thing always being available that means there’s never any urgency.

You may have heard of hyperbolic discounting from behavioral economics: people will generally disproportionally, i.e. hyperbolically, discount the value of something the farther off it is. The average person judges $15 now as equivalent to $30 in 3-months (an annual rate of return of 277%!).

It’s always worth gathering information, nurturing other projects, and putting together some backup plans. You’ll need to define what success means to you for each of them, because you won’t make overnight progress; instead, you’re best served picking projects that you can learn critical lessons from, even if you fail

“… there are about 25 billion car trips per year, and with some 27 million cars, this suggests an average of just under 18 trips per car every week. Since the duration of the average car trip is about 20 minutes, the typical car is only on the move for 6 hours in the week: for the remaining 162 hours it is stationary – parked.”

Extractive summarization may be regarded as acontextual bandit as follows. Each document is acontext, and each ordered subset of a document’ssentences is a different action

andit is a decision-making formal-ization in which an agent repeatedly chooses oneof several actions, and receives a reward based onthis choice.

BanditSum a hierarchical bi-LSTM

increasing body of research analytically exploresthe consequences of the research impact agenda on academic work,including the risks posed to research quality (Chubb and Reed2018), prioritising of short-term impacts rather than more concep-tual impacts (Greenhalgh and Fahy 2015; Meagher and Martin2017), ethical risks (Smith and Stewart 2017), and a focus on indi-vidual academics rather than on the broader context of research-based policy change (Dunlop 2018)

Unsur-prisingly, therefore, existing research documents various ways in which REF impact has becomeembedded within university governance, including via the broadening of career progression criteria(Bandola-Gill 2019)

RANDreport that had been commissioned by HEFCE (Grant et al. 2010)

while there aregroups potentially benefiting from the case studies relating to their field of research (egwriters benefiting from studies in Panel D, engineers benefiting from studies in PanelB), there are mentions of these potential beneficiaries across all the panels

With the benefit of hindsight, our analysis would have been much easierif the case studies had greater structure and used standardized definitions. Giventhat the case studies spanned a 20-year period, organization names have changed inthat time and keyword searches were not sophisticated enough to capture some keyinformation.

Topic modelling was used to determine common topics across the wholecorpus. Sixty-five topics were found (of which 60 were used) using theApache Mallet Toolkit Latent Dirichlet Allocation (LDA) algorithm.

any effect on, change or benefit to the economy, society,culture, public policy or services, health, the environment or quality of life, beyondacademia’ (REF, 2011).

A blog post is a very long and complex search query to find fascinating people and make them route interesting stuff to your inbox.

Research funders and providers are having to compete with other public services, and,as such, must be able to advocate the need for funding of research. Leaders within thesector must have compelling arguments to ‘make the case’ for research. For example,the Research Councils each publish an annual impact report which describe the waysin which they are maximising the impacts of their investments. These reports includeillustrations of how their research and training has made a contribution to the economyand society.10 The analysis of Researchfish and other similar data can support thedevelopment of these cases

Research outputs (and outcomes and impact) are gathered through a ‘questionset’ developed by funding institutions through a consultative process. This set of16 questions contains 175 sub-questions as illustrated in Figure 3 (the full set ofquestions are available in Annex A). A researcher, or one of their delegates, can add,edit and delete entries, and crucially, attribute entries to research grants and awards

The term ‘impact’ is currently used widely in research, especially with the inclusion ofnon-academic impact as part of the latest Research Excellence Framework (REF)

look at the economicimpact of research – taking an area of research(often cardiovascular disease), calculating thetotal investment in research and comparing it tothe total payback in terms of monetarised healthbenefit and other economic effects.

However, knowledge production isnormally only an intermediate aim: the ultimateobjective of most medical research is to improvehealth and prosperity.

Much broad and shallow evaluation is based onbibliometrics (examining the quality of researchpublications) to assess the amount and quality ofknowledge produced

Matthew Hindman, in his book "The Internet Trap" <http://assets.press.princeton.edu/chapters/s13236.pdf>, notes that most research on the internet has focused on its supposedly decentralized nature, leaving us with little language to really grapple with the concentrated, oligopolistic state of today's online economy, where the vast majority of attention and revenue accrue to a tiny number of companies

Nim in Action book

This isn't a highly scientific post full of esoteric details and language feature matrices. It's about making the best choice for what I can be the most productive in for my target market and product requirements.

Annotations are the first step of getting useful insights into my notes. This makes it a prerequisite to be able to capture annotations in my note making tool Obsidian, otherwise Hypothes.is is just another silo you’re wasting time on. Luckily h. isn’t meant as a silo and has an API. Using the API and the Hypothes.is-to-Obsidian plugin all my annotations are available to me locally.

In the same category of integrating h. into my pkm workflows, falls the interaction between h. and Zotero, especially now that Zotero has its own storage of annotations of PDFs in my library. It might be of interest to be able to share those annotations, for a more complete overview of what I’m annotating. Either directly from Zotero, or by way of my notes in Obsidian (Zotero annotatins end up there in the end)

Whatever your thing is, make the thing you wish you had found when you were learning. Don’t judge your results by “claps” or retweets or stars or upvotes - just talk to yourself from 3 months ago

a habit of creating learning exhaust:

I love the IndieWeb and its tools, but it has always bothered me that at some point they basically require you to have a webdevelopment background.

First, to experiment personally with AP itself, and if possible with the less known Activities that AP could support, e.g. travel and check-ins. This as an extension of my personal site in areas that WordPress, OPML and RSS currently can’t provide to me. This increases my own agency, by adding affordances to my site. This in time may mean I won’t be hosting or self-hosting my personal Mastodon instance. (See my current fediverse activities)

For example, the design pattern A Place to Wait asks that we create comfortable accommodation and ambient activity whenever someone needs to wait; benches, cafes, reading rooms, miniature playgrounds, three-reel slot machines (if we happen to be in the Las Vegas airport). This solves the problem of huddles of people awkwardly hovering in liminal space; near doorways, taking up sidewalks, anxiously waiting for delayed flights or dental operations or immigration investigations without anything to distract them from uncertain fates.

Misleading Templates There is no consistent re-lation between the performance of models trainedwith templates that are moderately misleading (e.g.{premise} Can that be paraphrasedas "{hypothesis}"?) vs. templates that areextremely misleading (e.g., {premise} Isthis a sports news? {hypothesis}).T0 (both 3B and 11B) perform better givenmisleading-moderate (Figure 3), ALBERT andT5 3B perform better given misleading-extreme(Appendices E and G.4), whereas T5 11B andGPT-3 perform comparably on both sets (Figure 2;also see Table 2 for a summary of statisticalsignificances.) Despite a lack of pattern between

Insum, notwithstanding prompt-based models’impressive improvement, we find evidence ofserious limitations that question the degree towhich such improvement is derived from mod-els understanding task instructions in waysanalogous to humans’ use of task instructions.

Suppose a human is given two sentences: “Noweapons of mass destruction found in Iraq yet.”and “Weapons of mass destruction found in Iraq.”They are then asked to respond 0 or 1 and receive areward if they are correct. In this setup, they wouldlikely need a large number of trials and errors be-fore figuring out what they are really being re-warded to do. This setup is akin to the pretrain-and-fine-tune setup which has dominated NLP in recentyears, in which models are asked to classify a sen-tence representation (e.g., a CLS token) into some

Antibiotic resistance has become a growingworldwide concern as new resistance mech-anisms are emerging and spreading globally,and thus detecting and collecting the cause– Antibiotic Resistance Genes (ARGs), havebeen more critical than ever. In this work,we aim to automate the curation of ARGs byextracting ARG-related assertive statementsfrom scientific papers. To support the researchtowards this direction, we build SCIARG, anew benchmark dataset containing 2,000 man-ually annotated statements as the evaluationset and 12,516 silver-standard training state-ments that are automatically created from sci-entific papers by a set of rules. To set upthe baseline performance on SCIARG, weexploit three state-of-the-art neural architec-tures based on pre-trained language modelsand prompt tuning, and further ensemble themto attain the highest 77.0% F-score. To the bestof our knowledge, we are the first to leveragenatural language processing techniques to cu-rate all validated ARGs from scientific papers.Both the code and data are publicly availableat https://github.com/VT-NLP/SciARG.

Our annotators achieve thehighest precision with OntoNotes, suggesting thatmost of the entities identified by crowdworkers arecorrect for this dataset.

an algorithm with high precision on LitBank orOntoNotes would miss a huge percentage of rele-vant mentions and entities on other datasets (con-straining our analysis)

Procedure: We first launch an annotation tutorial(paid $4.50) and recruit the annotators on the AMTplatform.9 At the end of the tutorial, each annotatoris asked to annotate a short passage (around 150words). Only annotators with a B3 score (Bagga

Annotation structure: Two annotation ap-proaches are prominent in the literature: (1) a localpairwise approach, annotators are shown a pairof mentions and asked whether they refer to thesame entity (Hladká et al., 2009; Chamberlain et al.,2016a; Li et al., 2020; Ravenscroft et al., 2021),which is time-consuming; or (2) a cluster-basedapproach (Reiter, 2018; Oberle, 2018; Bornsteinet al., 2020), in which annotators group all men-tions of the same entity into a single cluster. InezCoref we use the latter approach, which can befaster but requires the UI to support more complexactions for creating and editing cluster structures.

owever, these datasets vary widelyin their definitions of coreference (expressed viaannotation guidelines), resulting in inconsistent an-notations both within and across domains and lan-guages. For instance, as shown in Figure 1, whileARRAU (Uryupina et al., 2019) treats generic pro-nouns as non-referring, OntoNotes chooses not tomark them at all

Specifically, our work investigates the quality ofcrowdsourced coreference annotations when anno-tators are taught only simple coreference cases thatare treated uniformly across existing datasets (e.g.,pronouns). By providing only these simple cases,we are able to teach the annotators the concept ofcoreference, while allowing them to freely interpretcases treated differently across the existing datasets.This setup allows us to identify cases where ourannotators disagree among each other, but moreimportantly cases where they unanimously agreewith each other but disagree with the expert, thussuggesting cases that should be revisited by theresearch community when curating future unifiedannotation guidelines

this work, we developa crowdsourcing-friendly coreference annota-tion methodology, ezCoref, consisting of anannotation tool and an interactive tutorial. Weuse ezCoref to re-annotate 240 passages fromseven existing English coreference datasets(spanning fiction, news, and multiple other do-mains) while teaching annotators only casesthat are treated similarly across these datasets

One example could be putting all files into an Amazon S3 bucket. It’s versatile, cheap and integrates with many technologies. If you are using Redshift for your data warehouse, it has great integration with that too.

n recent years, the neural network based topic modelshave been proposed for many NLP tasks, such as infor-mation retrieval [11], aspect extraction [12] and sentimentclassification [13]. The basic idea is to construct a neuralnetwork which aims to approximate the topic-word distri-bution in probabilistic topic models. Additional constraints,such as incorporating prior distribution [14], enforcing di-versity among topics [15] or encouraging topic sparsity [16],have been explored for neural topic model learning andproved effective.

e argue that mutual learningwould benefit sentiment classification since it enriches theinformation required for the training of the sentiment clas-sifier (e.g., when the word “incredible” is used to describe“acting” or “movie”, the polarity should be positive)

. However, such a framework is not applicablehere since the learned latent topic representations in topicmodels can not be shared directly with word or sentencerepresentations learned in classifiers, due to their differentinherent meanings

“The metaphor is that the machine understands what I’m saying and so I’m going to interpret the machine’s responses in that context.”

Elicit is really impressive. It searches academic papers, providing summary abstracts as well as structured analyses of papers. For example, it tries to identify the outcomes analysed in the paper or the conflicts of interest of the authors, as well as easily tracks citations. (See a similar search on “technology transitions”. Log in required.)

I only know a handful of people directly making money from blogging (via ads, subscriptions etc) but I know many more who: Got a better career because of blogging (new job, better pay etc) Negotiated better contracts (e.g. with a publisher or platform) because they had “an audience” Sold their own courses / ebooks / books / merchandise / music Blogging is this kind of engine that opens up economic opportunity and advantage. Being visible in the networked economy has real value.

I’ve been using this phrase “the next most useful thing” as a guiding light for my consulting work - I’m obsessed with being useful not just right. I’ve always rejected the fancy presentation in favor of the next most useful thing, and I simply took my eye off the ball with this one. I’m not even sure the client views this project as a real disappointment, there was still some value in it, but I’m mad at myself personally for this one. A good reminder not to take your eye off the ball. And to push your clients beyond what they tell you the right answer is.

Being self-employed feels a bit like being on an extended road trip. Untethered and free, but lonely and unsupported too. Ultimate freedoms combined with shallow roots.

The actual reward state is not one where you're lazing around doing nothing. It's one where you're keeping busy, where you're doing things that stimulate you, and where you're resting only a fraction of the time. The preferred ground state is not one where you have no activity to partake in, it's one where you're managing the streams of activity precisely, and moving through them at the right pace: not too fast, but also not too slow. For that would be boring

the work that needs to be done is not a finite list of tasks, it is a neverending stream. Clothes are always getting worn down, food is always getting eaten, code is always in motion. The goal is not to finish all the work before you; for that is impossible. The goal is simply to move through the work. Instead of struggling to reach the end of the stream, simply focus on moving along it.

It took me a while to grok where dbt comes in the stack but now that I (think) I have it, it makes a lot of sense. I can also see why, with my background, I had trouble doing so. Just as Apache Kafka isn’t easily explained as simply another database, another message queue, etc, dbt isn’t just another Informatica, another Oracle Data Integrator. It’s not about ETL or ELT - it’s about T alone. With that understood, things slot into place. This isn’t just my take on it either - dbt themselves call it out on their blog:

Of course, despite what the "data is the new oil" vendors told you back in the day, you can’t just chuck raw data in and assume that magic will happen on it, but that’s a rant for another day ;-)

Working with the raw data has lots of benefits, since at the point of ingest you don’t know all of the possible uses for the data. If you rationalise that data down to just the set of fields and/or aggregate it up to fit just a specific use case then you lose the fidelity of the data that could be useful elsewhere. This is one of the premises and benefits of a data lake done well.

But - as the overall network has grown exponentially the network topology has changed. Digg, Reddit, Hacker News etc all still exist but the audience you can reach with a “homepage” hit there has become much smaller relative to the overall size of the network. And getting a homepage hit there is harder than ever because the volume of content has increased exponentially

Many people report writers block with blogs, particularly after a big successful post, because it’s almost impossible to consistently pump out bangers.