All Courses - Page 85 of 394 - Analytics Campus

Can LLMs Replicate a PNAS Paper? (Half 2)

Econometrics

Dr. Mike

February 6, 2026

Can LLMs Replicate a PNAS Paper? (Half 2)

I used gpt-4o-mini to duplicate the textual content classification from Card et al.’s PNAS paper on 140 years of immigration rhetoric. Right here’s what occurred:

Complete Value of this Train: $10.99
Complete Time: 1 hour arrange (yesterday’s video), 2.6 hours (we anticipated 24), 1 hour of research in the present day = 4.6 hours.
Settlement with authentic classifier: 69%
Key discovering: The polarization story is strong—each classifiers present Democrats and Republicans diverging sharply because the Nineteen Seventies

Backside line: LLMs can considerably replicate skilled classifier outcomes at a fraction of the fee. The substantive conclusions survive. For those who’re doing exploratory textual content evaluation with out sources to fine-tune a customized mannequin, that is now a viable path.

The catch: The LLM is extra “cautious”—when unsure, it calls issues NEUTRAL as a substitute of choosing a facet. Direct PRO↔ANTI disagreements are uncommon (solely 4%).

The video is under, and a phrase of thanks to all paying subscribers and people who have gifted the substack and shared the posts. Thanks! This substack is a labor of affection, and I actually take pleasure in sharing what I’m studying about utilizing Claude Code for quantitative social scientific analysis. So thanks everybody on your help.

Now for the small print.

In Half 1 (hyperlink under), I described an experiment: may Claude Code assist me replicate the textual content classification from Card et al.’s PNAS paper on 140 years of immigration rhetoric? And in yesterday’s publish, Claude Code internet crawled to seek out the replication package deal, organized my directories, got here up with a plan, cracked open the information, submitted json information in batch to OpenAI to have the gpt-4o-mini do a one-shot evaluation of 304,995 speeches able to classify. I used a batch job which is 50% cheaper than doing it one request at a time, plus it’s 305k speeches so you possibly can’t realistically try this anyway. However there was a number of uncertainty as to what would occur.

Initially, yesterday, Claude Code informed me that it could take 24 hours for the evaluation to be accomplished, however it truly took solely 2.6 hours and doubtless much less as a result of I needed to resubmit a few batches that had failed and solely found that was the case at round 2.5 hours. So I wasn’t positive how lengthy it had been finished is my level once I resent in these final two batches. Anyway, it’s finished, the outcomes are in, and I filmed myself analyzing it, in addition to making a deck (which I am going via within the video).

Right here’s what we discovered.

First, simply so this has its personal part and like I stated, once I submitted the batch job, I anticipated to attend. And that’s as a result of OpenAI’s documentation says batch jobs can take “as much as 24 hours.” I believed initially that when Claude Code informed me, due to this fact, it could possible take 24 hours, it was as a result of he did some again of the envelope calculation, however no. he was simply counting on the time stamp. I figured I’d verify again the following morning. However once I checked the time stamps, it had taken round 2 1/2 hours to do it.

That’s not a typo. Two and a half hours to categorise almost 300,000 speeches.

For context: the unique paper’s authors needed to fine-tune a RoBERTa mannequin on 7,626 human annotations, then run inference on their complete corpus. That’s weeks of labor in case you embody the annotation time, possibly extra relying on what number of RAs they most likely needed to rent to try this annotation.

However that’s not what we did. I truly initially thought in actual fact that we had been utilizing the 7,626 human annotations on this train, however we weren’t. Moderately we did a zero-shot replication . For eleven {dollars}.

Let me provide the headline quantity first, then we’ll unpack what it means.

gpt-4o-mini agreed with the unique RoBERTa classifier on 69% of speeches.

Is that good? It is determined by your baseline.

Right here’s what helped me calibrate: the unique paper experiences that human annotators agreed with one another at Krippendorff’s α = 0.48. That’s thought-about “average” settlement in content material evaluation. People themselves usually disagreed about whether or not a speech was pro-immigration, anti-immigration, or impartial.

If even people can’t reliably agree, how a lot ought to we count on two totally different ML programs to agree?

69% isn’t good. But it surely’s much better than probability (which might be ~33% for a three-class downside). And it’s within the ballpark of what you’d count on given the underlying ambiguity of the duty.

Right here’s the complete image:

Let me stroll you thru the best way to learn this.

Every row is an authentic RoBERTa label. Every column is what gpt-4o-mini categorized it as. The odds are row percentages—of all speeches RoBERTa labeled as X, what fraction did the LLM label as Y? They need to sum to 100% studying let to proper as a result of they’re treating the unique classification because the “pattern area” so to talk after which breaking apart the share of complete new classification into the three totally different classes all of that are mutually unique.

The diagonal is settlement between outdated and new classifications. NEUTRAL has the best settlement (85%). PRO has first rate settlement (63%). ANTI has the bottom (51%).

The off-diagonal tells you about systematic disagreements. And right here’s the attention-grabbing sample:

When gpt-4o-mini disagrees with RoBERTa, it nearly at all times strikes towards NEUTRAL.

Of speeches RoBERTa known as PRO: 63% the LLM additionally known as PRO, however 33% the LLM known as NEUTRAL
Of speeches RoBERTa known as ANTI: 51% the LLM additionally known as ANTI, however 44% the LLM known as NEUTRAL

The LLM is extra cautious. When it’s unsure, it hedges towards the center class.

Right here’s what anxious me entering into: what if the 2 classifiers basically disagreed about tone? What if speeches RoBERTa known as PRO-IMMIGRATION, the LLM known as ANTI?

That might be a major problem. It will imply the classifiers have incompatible understandings of political language.

The info is reassuring:

PRO → ANTI: 3.7%
ANTI → PRO: 4.9%

Direct polarity flips are uncommon. When the classifiers disagree, they normally disagree about whether or not one thing is impartial vs. opinionated—not about which path the opinion factors.

This issues for the substantive findings. For those who’re utilizing these classifications to trace partisan polarization over time, what you actually care about is whether or not a speech is clearly professional, clearly anti, or ambiguous. The 2 classifiers largely agree on that construction.

Let me present you Determine 1 from the unique paper:

That is the important thing discovering: Democrats and Republicans have polarized sharply because the Nineteen Seventies. Democrats use more and more optimistic language about immigrants. Republicans use language as damaging because the Nineteen Twenties quota period.

Now right here’s the identical determine utilizing our LLM classifications:

The polarization story is strong.

Each classifiers present the identical fundamental sample:

Partisan divergence beginning within the Nineteen Seventies
Democrats trending extra optimistic
Republicans trending extra damaging
The hole widening over time

The magnitudes differ—the LLM model reveals extra volatility, particularly within the early interval. However the form of the story is identical.

That is an important discovering for anybody questioning whether or not LLM-based classification can replicate conventional NLP outcomes. The qualitative conclusions maintain up. A researcher utilizing gpt-4o-mini would attain the identical substantive interpretation as the unique authors.

I used to be curious whether or not the patterns seemed totally different for my residence state, Texas, so I ran the identical evaluation restricted to Texas congressional speeches:

It’s noisier, and I believe there should not be many Republican Texas congressional speeches within the information earlier than the second half of the twentieth century. And when there’s any speeches by Texans, they’re Democrats and curiously, they’re damaging in the direction of immigration. However that modifications after the warfare and Democrats in Texas grow to be more and more extra optimistic. There’s an enormous pro-immigration Texas spike within the mid-Nineteen Seventies, however the identical polarization hole occurs with Texas as occurs nationally beginning round 1980.

So, the polarization story holds, although with extra noise (smaller pattern). And Texans being extra damaging in the direction of immigration seems to be a little bit of a posh historic story that I’d like to dig into extra however most likely received’t.

The unique paper additionally analyzed how tone various by the nation of origin being mentioned—Mexico, China, Italy, and many others. We replicated that:

Now evaluate to the unique:

Discover the way it isn’t merely that the general developments match between the unique Roberta classification. Moderately, it’s even the ordering that’s the identical. Discover how within the authentic, it was Italy>China>Mexico, which is identical ordering because the LLM classification. Different findings:

China: Most damaging within the early interval (Chinese language Exclusion Period), bettering after the Nineteen Sixties
Mexico: Comparatively secure traditionally, with a dip within the trendy period (border politics)
Italy: Most optimistic general, particularly after WWII (”mannequin minority” framing)

And the LLM captured the country-specific patterns with none fine-tuning. It understood that speeches mentioning “Chinese language” in 1890 had totally different connotations than speeches mentioning “Chinese language” in 1990.

Let me be clear about one thing: this was not an apples-to-apples comparability.

RoBERTa (the unique):

Fantastic-tuned on 7,626 human-labeled examples
Skilled particularly for this job
Optimized to match annotator judgments
~65% accuracy on held-out information

gpt-4o-mini (our replication):

Zero-shot (no task-specific coaching)
Common language mannequin
Inferring the duty from a immediate
No publicity to the unique annotations

The RoBERTa mannequin was taught what pro-immigration and anti-immigration imply within the context of this particular corpus. gpt-4o-mini needed to determine it out from first ideas.

Provided that asymmetry, 69% settlement is definitely spectacular. The LLM introduced real language understanding to the duty, not simply sample matching on the particular options the RoBERTa mannequin discovered.

Greatest I can inform, what’s going on right here is that after they disagree, it’s at these “marginal speeches” that had been proper on the sting between anti and impartial, or professional and impartial. And if you reclassify them, you’re largely pushing the marginal speeches into the impartial territory. Which I believe most likely means they’re roughly random within the information, and as such the sign coming via from the professional and anti within the authentic classification is fairly robust leaving the general patterns intact.

LLMs can considerably replicate skilled classifier outcomes.

69% settlement isn’t good, however it’s adequate that the substantive findings survive. For those who’re doing exploratory textual content evaluation and don’t have sources to fine-tune a customized mannequin, gpt-4o-mini is a viable possibility.

The polarization discovering from Card et al. is strong.

Each classifiers inform the identical story about partisan divergence. This will increase my confidence within the authentic paper’s conclusions. The discovering isn’t an artifact of their particular mannequin.

LLMs are extra conservative.

gpt-4o-mini most popular NEUTRAL when unsure. That is most likely good conduct—it’s higher to say “I don’t know” than to confidently guess improper. But it surely means LLM-based classification might underestimate the proportion of clearly opinionated textual content.

Direct polarity flips are uncommon.

The classifiers hardly ever known as the identical speech each PRO and ANTI. After they disagreed, it was normally about whether or not one thing was impartial vs. opinionated. That is reassuring for anybody utilizing these instruments for sentiment evaluation.

LLM was cheaper

And maybe an important factor — the fee. One thing like 4 hours, begin to end, and solely $11. The onerous half was due to this fact most likely in getting the information — however I wager you that too is a quest we may undertake at appreciable velocity given it’s totally potential these can be found to us such that Claude Code would’ve crawled and scraped the speeches, or they’re saved in some warehouse someplace. Although possibly this crew themselves transcribed all 305k of them; I’ll have to verify extra carefully to see. I believe nonetheless that the true bottleneck goes to be on the margins of getting the information and any DUAs that don’t can help you use an AI agent for evaluation, however I nonetheless consider we’ll see artistic methods the place researchers accommodate their DUAs to the letter however we nonetheless use AI brokers to hurry issues up. As an illustration, simply the creating of the audited code ex ante earlier than you enter some closed off location is itself alone going to be potential.

Past the substantive findings, this experiment taught me issues about utilizing LLMs for analysis:

The Batch API is underrated. Most researchers I do know use ChatGPT interactively. Possibly, they name the API one request at a time. However that’s an enormous possibly. Utilizing the Batch API is cheaper (50% off), handles giant jobs gracefully, and completed method quicker than promised. For those who’re doing any type of textual content evaluation at scale, study the Batch API. However you may also have Claude Code do it for you and clarify it to you in the best way your mind discovered alongside the best way.
Referee 2 was important. Having a separate Claude occasion overview the code caught actual bugs: edge instances in label normalization, lacking metrics (Cohen’s Kappa), considerations about immediate design. The code that ran was higher than the code I’d have submitted with out overview.
Not sticking to my deliberate workflow. However on the similar time, yesterday I inadvertently used referee 2 from inside the similar context window which as I stated violates the whole level of utilizing referee 2. We don’t need college students grading their very own exams. And we don’t need audits to be finished by the very Claude Code manifestation that wrote the code. So one way or the other there’s a psychological factor in regards to the referee 2 workflow that’s nonetheless not clicking for me, and I’m unsure what it’s. I don’t know if it’s as a result of I’m filming myself working, or what, however the level is, you received’t get an error if you audit your code utilizing referee2 from the wrong Claude Code context window. Which signifies that I can audit my code incorrectly and by no means know. I solely know as a result of a reader/viewer caught it and informed me, which has given me pause about how workflows will work now.
$11 will not be some huge cash. The fee barrier for this sort of analysis is now trivially low. The barrier is figuring out the best way to set it up—which is why I wrote this.
2.6 hours will not be a number of time. I anticipated to attend in a single day. I barely completed dinner. For those who’re iterating on immediate designs or testing totally different fashions, you possibly can run a number of experiments in a day.

Let me put the sensible numbers in a single place:

That catchphrase within the caption — Lower than a Netflix subscription to duplicate a PNAS paper — was Claude Code’s personal astonishment at this entire factor!

Three years in the past, in case you wished to categorise 300,000 textual content paperwork, you had two choices: rent human annotators (costly, gradual) or practice a customized NLP mannequin (requires experience, compute, labeled information).

Now there’s a 3rd possibility: ship your paperwork to a general-purpose LLM and get ends in just a few hours for pocket change. And use Claude Code (or Codex or one thing like them each) that will help you do it so that you don’t mess it up.

The outcomes received’t be equivalent to a fine-tuned mannequin, however that’s the purpose. It’s a unique NLP classification. This train wasn’t about making an attempt to talk into that, although. I actually didn’t know what I used to be going to seek out. I simply wished to indicate that Claude Code pulled all this collectively from the replication package deal it discovered itself on-line, set it up and submitted the batches to openai. All for less than $11 and round 4 1/2 hours time begin to end, together with making the cool decks to show me it. For a lot of analysis questions, that’s adequate for simply gaining fast and lasting insights.

This modifications who can do computational textual content evaluation. You don’t want a machine studying background. You don’t want GPU clusters. You don’t want a finances for annotation.

You want a analysis query, some textual content information, and eleven {dollars}.

Authentic paper:
Card, D., Chang, S., Becker, C., Mendelsohn, J., Voigt, R., Boustan, L., Abramitzky, R., & Jurafsky, D. (2022). Computational evaluation of 140 years of US political speeches reveals extra optimistic however more and more polarized framing of immigration. PNAS, 119(31), e2120510119.

Replication information:
github.com/dallascard/us-immigration-speeches

Our replication:

That is Half 2 of a two-part collection on utilizing LLMs for analysis replication. Half 1 lined the setup; this half lined the outcomes. For those who’re excited about making an attempt this your self, the important thing instruments are Claude Code for orchestration, the OpenAI Batch API for scale, and a wholesome dose of methodological skepticism.

How Associa transforms doc classification with the GenAI IDP Accelerator and Amazon Bedrock

Machine Learning

Dr. Mike

February 6, 2026

How Associa transforms doc classification with the GenAI IDP Accelerator and Amazon Bedrock

This can be a visitor publish co-written with David Meredith and Josh Zacharias from Associa.

Associa, North America’s largest group administration firm, oversees roughly 7.5 million householders with 15,000 workers throughout greater than 300 department places of work. The corporate manages roughly 48 million paperwork throughout 26 TB of knowledge, however their present doc administration system lacks environment friendly automated classification capabilities, making it troublesome to arrange and retrieve paperwork throughout a number of doc varieties. Each day, workers spend numerous hours manually categorizing and organizing incoming paperwork—a time-consuming, error-prone course of that creates bottlenecks in operational effectivity and probably ends in operational delays and lowered productiveness.

Associa collaborated with the AWS Generative AI Innovation Heart to construct a generative AI-powered doc classification system aligning with Associa’s long-term imaginative and prescient of utilizing generative AI to realize operational efficiencies in doc administration. The answer routinely categorizes incoming paperwork with excessive accuracy, processes paperwork effectively, and supplies substantial value financial savings whereas sustaining operational excellence. The doc classification system, developed utilizing the Generative AI Clever Doc Processing (GenAI IDP) Accelerator, is designed to combine seamlessly into present workflows. It revolutionizes how workers work together with doc administration programs by lowering the time spent on guide classification duties.

This publish discusses how Associa is utilizing Amazon Bedrock to routinely classify their paperwork and to assist improve worker productiveness.

Resolution overview

The GenAI IDP Accelerator is a cloud-based doc processing answer constructed on AWS that routinely extracts and organizes info from varied doc varieties. The system makes use of OCR know-how and generative AI to transform unstructured paperwork into structured, usable information whereas scaling seamlessly to deal with excessive doc volumes.

The accelerator is constructed with a versatile, modular design utilizing AWS CloudFormation templates that may deal with various kinds of doc processing whereas sharing core infrastructure for job administration, progress monitoring, and system monitoring. The accelerator helps three processing patterns. We use Sample 2 for this answer utilizing OCR (Amazon Textract) and classification (Amazon Bedrock). The next diagram illustrates this structure.

We optimized the doc classification workflow by evaluating three key elements:

Immediate enter – Full PDF doc (all pages) vs. first web page solely
Immediate design – Multimodal prompting with OCR information (utilizing the Amazon Textract analyze_document_layout) vs. doc picture solely
Mannequin selection – Amazon Nova Lite, Amazon Nova Professional, Amazon Nova Premier, and Anthropic’s Claude Sonnet 4 on Amazon Bedrock

This complete analysis framework helped us determine the configuration that delivers the best accuracy whereas minimizing processing inference prices for Associa’s particular doc varieties and operational necessities. The analysis dataset consists of 465 PDF paperwork throughout eight distinct doc varieties. The dataset contains some samples recognized as draft paperwork or e-mail correspondences. These samples are categorized as doc kind Unknown on account of inadequate classification standards. The distribution of doc varieties throughout courses is unbalanced, starting from 6 samples for Insurance policies and Resolutions to 155 samples for Minutes.

Analysis: Immediate enter

We began our preliminary analysis utilizing full PDF paperwork, the place all pages of a PDF have been used as enter to the immediate for classification. The next desk exhibits the accuracy for full PDF classification utilizing Amazon Nova Professional with OCR and picture. We noticed a median classification accuracy of 91% contemplating the completely different doc varieties with a median value of 1.10 cents per doc.

Doc Kind	Variety of Samples	Variety of Samples Categorized Accurately	Classification Accuracy	Classification Value (in Cents)
Bylaws	46	42	91%	1.52c
CCR Declarations	22	19	86%	1.55c
Certificates of Insurance coverage	74	74	100%	1.49c
Contracts	71	66	93%	1.48c
Minutes	155	147	95%	1.47c
Plat Map	21	20	95%	1.45c
Insurance policies and Resolutions	6	5	83%	0.35c
Guidelines and Laws	50	44	88%	0.36c
Unknown	20	8	40%	0.24c
General	465	425	91%	1.10c

Utilizing full PDF for doc classification demonstrates an accuracy of 100% for Certificates of Insurance coverage and 95% for Minutes. The system appropriately categorized 425 out of 465 paperwork. Nonetheless, for the Unknown doc kind, it achieved solely 40% accuracy, appropriately classifying simply 8 out of 20 paperwork.

Subsequent, we experimented with utilizing solely the primary web page of a PDF doc for classification, as proven within the following desk. This method improved general accuracy from 91% to 95% with 443 out of 465 paperwork categorized appropriately whereas lowering classification value per doc from 1.10 cents to 0.55 cents.

Doc Kind	Variety of Samples	Variety of Samples Categorized Accurately	Classification Accuracy	Classification Value (in Cents)
Bylaws	46	44	96%	0.55c
CCR Declarations	22	21	95%	0.55c
Certificates of Insurance coverage	74	74	100%	0.59c
Contracts	71	64	90%	0.56c
Minutes	155	153	99%	0.55c
Plat Map	21	17	81%	0.56c
Insurance policies and Resolutions	6	4	67%	0.57c
Guidelines and Laws	50	49	98%	0.56c
Unknown	20	17	85%	0.55c
General	465	443	95%	0.55c

Other than improved accuracy and lowered value, the first-page-only method considerably improved Unknown doc classification accuracy from 40% to 85%. First pages usually include essentially the most distinctive doc options, whereas later pages in drafts or e-mail threads can introduce noise that confuses the classifier. Mixed with quicker processing speeds and decrease infrastructure prices, we chosen the first-page-only method for the following evaluations.

Analysis: Immediate design

Subsequent, we experimented on immediate design to judge whether or not OCR information is important for doc classification or simply utilizing the doc picture is adequate. We evaluated by eradicating the OCR textual content extraction information from the immediate and solely utilizing the picture in a multimodal immediate. This method removes the Amazon Textract prices and depends totally on the mannequin’s understanding of visible options. The next desk exhibits the accuracy for first-page-only classification utilizing Amazon Nova Professional with solely picture.

Doc Kind	Variety of Samples	Variety of Samples Categorized Accurately	Classification Accuracy	Classification Value (in Cents)
Bylaws	46	45	98%	0.19c
CCR Declarations	22	20	91%	0.19c
Certificates of Insurance coverage	74	74	100%	0.18c
Contracts	71	63	89%	0.18c
Minutes	155	151	97%	0.18c
Plat Map	21	18	86%	0.19c
Insurance policies and Resolutions	6	4	67%	0.18c
Guidelines and Laws	50	48	96%	0.18c
Unknown	20	10	50%	0.18c
General	465	433	93%	0.18c

The image-only classification method demonstrates comparable points as the total PDF classification method. Though this methodology achieves an general accuracy of 93%, for Unknown doc varieties, it may classify solely 10 out of 20 paperwork appropriately with 50% accuracy. The next desk summarizes our analysis of an image-only method.

	General Classification Accuracy (All Doc Sorts, Together with Unknown)	Classification Accuracy (Doc Kind: Unknown)	Classification Value (in Cents)
First web page solely classification (OCR + Picture)	95%	85%	0.55c
First web page solely classification (Solely Picture)	93%	50%	0.18c

The image-only method removes OCR prices however reduces general accuracy from 95% to 93% and Unknown doc accuracy from 85% to 50%. Correct Unknown doc classification is vital for downstream human evaluate and operational effectivity at Associa. We chosen the mixed OCR and picture method to keep up this functionality.

Analysis: Mannequin selection

Utilizing the optimum configuration of first-page-only classification with OCR and picture, we evaluated completely different fashions to determine an optimum stability of accuracy and price, as summarized within the following desk. We concentrate on general classification efficiency, classification of unknown paperwork, and per-document classification prices.

	General Classification Accuracy (All Doc Sorts, Together with Unknown)	Classification Accuracy (Doc Kind: Unknown)	Classification Value (in Cents)
Amazon Nova Professional	95%	85%	0.55c
Amazon Nova Lite	95%	50%	0.41c
Amazon Nova Premier	96%	90%	1.12c
Anthropic Claude Sonnet 4	95%	95%	1.21c

General classification accuracy ranged from 95–96% throughout the fashions, with variation in unknown doc kind efficiency. Certificates of Insurance coverage, Plat Map, and Minutes achieved 98–100% accuracy throughout the fashions. Anthropic’s Claude Sonnet 4 achieved the best unknown doc accuracy (95%), adopted by Amazon Nova Premier (90%) and Amazon Nova Professional (85%). Nonetheless, Anthropic’s Claude Sonnet 4 elevated classification value from 0.55 cents to 1.21 cents per doc. Amazon Nova Premier achieved one of the best general classification accuracy at 1.12 cents per doc. Contemplating the trade-offs between accuracy and price, we chosen Amazon Nova Professional because the optimum mannequin selection.

Conclusion

Associa constructed a generative AI-powered doc classification system utilizing Amazon Nova Professional on Amazon Bedrock that achieves 95% accuracy at a median value of 0.55 cents per doc. The GenAI IDP Accelerator facilitates dependable efficiency scaling to excessive quantity of paperwork throughout their branches. “The answer developed by AWS Generative AI Innovation Heart improves how our workers handle and manage paperwork, and we foresee important discount of guide effort in doc processing,” says Andrew Brock, President, Digital & Expertise Providers & Chief Info Officer at Associa. “The doc classification system supplies substantial value financial savings and operational enhancements, whereas sustaining our excessive accuracy requirements in serving residential communities.”

Seek advice from the GenAI IDP Accelerator GitHub repository for detailed examples and select Watch to remain knowledgeable on new releases. If you happen to’d prefer to work with the AWS GenAI Innovation Heart, please attain out to us or depart a remark.

Acknowledgements

We wish to thank Mike Henry, Bob Strahan, Marcelo Silva, and Mofijul Islam for his or her important contributions, strategic choices, and steering all through.

In regards to the authors

David Meredith is Director of Worker Software program Improvement at Associa. He oversees the efforts of the Associa staff to create software program for his or her 15,000 workers to make use of each day. He has virtually 20 years of expertise with software program within the residential property administration trade and lives within the Vancouver space of BC, Canada.

Josh Zacharias is a Software program Developer at Associa, the place he’s a lead engineer for the interior software program staff. His work contains architecting full stack options for varied departments within the firm in addition to empowering different builders to be extra environment friendly consultants in growing software program.

Monica Raj is a Deep Studying Architect on the AWS Generative AI Innovation Heart, the place she works with organizations throughout varied industries to develop AI options. Her work focuses on constructing and deploying agentic AI options, pure language processing, contact middle automation, and clever doc processing. Monica has in depth expertise in constructing scalable AI options for enterprise clients.

Tryambak Gangopadhyay is a Senior Utilized Scientist on the AWS Generative AI Innovation Heart, the place he collaborates with organizations throughout a various spectrum of industries. His function entails researching and growing generative AI options to handle essential enterprise challenges and speed up AI adoption. Previous to becoming a member of AWS, Tryambak accomplished his PhD at Iowa State College.

Nkechinyere Agu is an Utilized Scientist on the AWS Generative AI Innovation Heart, the place she works with organizations throughout varied industries to develop AI options. Her work focuses on growing multimodal AI options, agentic AI options, and pure language processing. Previous to becoming a member of AWS, Nkechinyere accomplished her PhD at Rensselaer Polytechnic Institute, Troy NY.

Naman Sharma is a Generative AI Strategist on the AWS Generative AI Innovation Heart, the place he collaborates with organizations to drive adoption of generative AI to unravel enterprise issues at scale. His work focuses on main clients from scoping, deploying, and scaling frontier options with the GenAIIC Technique and Utilized Science groups.

Yingwei Yu is an Utilized Science Supervisor on the Generative AI Innovation Heart, based mostly in Houston, Texas. With in depth expertise in utilized machine studying and generative AI, Yingwei leads the event of revolutionary options throughout varied industries.

Dwaragha Sivalingam is a Senior Options Architect specializing in generative AI at AWS, serving as a trusted advisor to clients on cloud transformation and AI technique. With eight AWS certifications, together with ML Specialty, he has helped clients in lots of industries, together with insurance coverage, telecom, utilities, engineering, development, and actual property. A machine studying fanatic, he balances his skilled life with household time, having fun with highway journeys, motion pictures, and drone images.

26 causes to look at the 2026 Winter Olympics in Milan

Technology

Dr. Mike

February 6, 2026

26 causes to look at the 2026 Winter Olympics in Milan

This story appeared in In the present day, Defined, a each day publication that helps you perceive essentially the most compelling information and tales of the day. Subscribe right here.

Vox loves the Olympics. Completely loves them. We briefly debated popping up a limited-run Winter Olympics publication this week, however then remembered that we’re not actually a sports activities website and don’t have any precise sports activities reporters on workers.

Irrespective of! The Olympics are enjoyable (and, sometimes, inspiring and heartbreaking and anxious and bizarre) as a result of they contact a variety of human pursuits, from geopolitics to local weather to movie star and tradition. And in that spirit, I’ve polled my colleagues and poked my head into a couple of planning classes to see what Olympics storylines everyone seems to be watching.

In at present’s version, we spherical up the attention-grabbing characters, open questions, and rising developments that would outline the Milan Cortina Video games, from the debut of skimo (fascinating! enjoyable!) to the sophisticated ethical calculus of rooting for Crew USA proper now (nuanced, uncomfortable, in brief: a bummer).

The entire shebang formally kicks off this afternoon at 2 pm ET, with a primetime broadcast at 8 pm. You’ll be able to completely wager that I and lots of others at Vox will probably be tuning in.

Are we the baddies? I believe it’s going to be fascinating to see how American followers react to, properly, being world villains. I believe it may very well be significantly attention-grabbing round Olympic hockey. In all probability essentially the most well-known second of American underdogness — not less than for the reason that American Revolution — was the 1980 Olympics hockey win. The Individuals and the Canadians are the highest rivals in Olympic hockey, and these Video games will probably be intense. We’re not plucky underdogs — the US group most likely has the second-best probability on the gold after Canada — and we is not going to be the gang favourite. What’s that going to really feel like? —Bryan Walsh

Skimo. I’m nonetheless rooting for the Winter Olympics so as to add two of my favourite sports activities — cross-country operating and cyclocross bike racing, each of that are nice within the snow — however for now, I’m loads excited in regards to the new sport we did get: skimo, or ski mountaineering. Because the title suggests, it entails summiting a climb (partially with skis, partially with out) after which tearing again down it. Athletes should transition between phases alongside the best way, triathlon-style, and it appears to be like thrilling. —Cameron Peters

Ilia Malinin. Twenty-one-year-old determine skater Ilia Malinin is competing in his first Olympics this 12 months, however he’s already damaged a ton of information. In 2022, he turned the primary (and nonetheless solely) skater to land a totally rotated quadruple axel in worldwide competitors, and in December, he landed a report seven quadruple jumps in a single program. In different phrases, Malinin is a mind-bogglingly aggressive, bodily skater — The Atlantic dubbed him “the person who broke physics” — and his applications will probably be actually enjoyable to look at once they kick off this weekend. —Caitlin Dewey

A determine skater’s tragic backstory. Maxim Naumov is a member of the US Olympic determine skating group whose dad and mom — world champions in skating themselves — have been among the many 67 folks killed when an American Airways airplane and a helicopter collided over the Potomac River in Washington, DC, final January. As a lot because the airplane crash affected DC, it was much more devastating for the determine skating neighborhood, which misplaced 28 dad and mom, skaters, alumni, volunteers, and extra. It’s onerous to not root for Naumov within the aftermath of such disappointment. —Libby Nelson

US immigration brokers. The Division of Homeland Safety usually has a presence on the Olympic Video games, offering safety companies and monitoring legal exercise. However the information that some ICE brokers would deploy to Milan has sparked outrage and protests in Italy. Milan’s governor known as ICE “a militia that kills” and mentioned “they don’t seem to be welcome” in his metropolis. —CD

Potential political demonstrations. Tons of of athletes competing in Milan hail from one of many world’s most politically unstable nations: the USA of America. As Vox’s Alex Abad-Santos has written, American athletes are certain to face questions about representing the nation of Donald Trump. It appears doubtless that some US Olympians, lots of whom are immigrants or the kids of immigrants, will communicate out towards ICE; I will probably be watching to see who does and what they are saying. Finally 12 months’s 4 Nations males’s hockey match, pro-MAGA and anti-Trump sentiment coloured the US-Canada remaining, with Trump himself calling the boys’s group and elevating the tempers of Canadian gamers and followers along with his calls to annex what he was calling “the 51st state.” —Seth Maxon

…and never simply from Crew USA. There are 4 Israelis competing at these Video games who could face protests or develop into fulcrums of controversy. And there are Ukrainians, “unaffiliated” athletes from Russia, a few Iranians, and even a Venezuelan cross-country skier competing. Whether or not any of them communicate up or are confronted about their nations’ turmoil will probably be price watching out for. —SM

Winter climate. Many Winter Olympic occasions hinge on splendid out of doors snow and ice circumstances which are changing into more durable to attain because the planet warms up. The IOC mentioned they’re now contemplating shifting the Winter Video games earlier within the 12 months to chase the chilly climate. Alternatively, among the early ski classes for the present Video games have been canceled as a consequence of an excessive amount of snow. —Umair Irfan

Milo and Tina. I don’t suppose any Olympic mascots will ever high Paris’s anthropomorphic hats, however maintain an eye fixed out for the scarf-wearing, short-haired weasels that organizers have known as the Video games’ “first overtly Gen Z” representatives. The open Gen Z-ness relates to their backstories (Milo and Tina are reportedly youngsters) and their design origins (Italy invited major and secondary faculty college students to submit mascot proposals). —CD

Alysa Liu. Not solely is 20-year-old determine skater Alysa Liu competing on the most important stage on this planet after a virtually two-year hiatus from the game, however Liu informed her coaches that she wished to design her personal costumes and skate to her alternative of music. Up to now, that’s concerned a number of Woman Gaga, a smiley piercing, and halo-striped hair. I can’t wait to see what she and the remainder of the US determine skating group accomplish. —Sydney Bergen

Macklin Celebrini. The time period “generational expertise” will get tossed round loads when hockey watchers discuss in regards to the 19-year-old ahead from Vancouver. He’s made a stagnant, struggling group (my group, the San Jose Sharks) aggressive as soon as once more. Celebrini is about to develop into the youngest participant to characterize Canada in a Winter Olympics that includes gamers from the Nationwide Hockey League — the place he’ll play alongside NHL icons like Sidney Crosby, Nathan MacKinnon, and Connor McDavid. Celebrini has been an enchanting participant to look at when a complete group depends on him, so I can’t think about what he can accomplish when he’s enjoying alongside his sport’s largest stars. —Christian Paz

Lindsey Vonn. The 41-year-old American skier fully ruptured her ACL in a fall final week, however has insisted that she’ll compete in her fifth Winter Video games regardless of the (extreme and debilitating!) damage. If she competes as deliberate, Vonn would be the oldest feminine Alpine skier in Olympic historical past. As an individual round Vonn’s age, and because the proprietor of a totally reconstructed ACL, I’m each vaguely horrified for Vonn and really invested in her competitors. —CD

Chloe Kim. The American snowboarding celebrity suffered a shoulder damage in early January that has prevented her from coaching within the weeks which have adopted. Fortunately, she has recovered sufficient to nonetheless compete. However is she nonetheless in high kind and in a position to contend for gold as anticipated? I hope so! Both manner, I’m certain NBC and Peacock are already getting ready their scripts about how a lot adversity she’s overcome to get again on the rostrum. —SM

Laila Edwards. The 22-year-old senior on the College of Wisconsin made historical past in Italy by simply being there: Edwards is the primary Black girl to play for the US ladies’s Olympic hockey group. And due to a profitable GoFundMe marketing campaign — which included a donation from the Kelce brothers, who’re from her hometown of Cleveland Heights, Ohio — Edwards’s dad and mom and different members of the family have made the journey to Milan as properly. —Esther Gim

Eileen Gu. A breakout star on the 2022 Olympics is again to compete in freestyle snowboarding. On the final Video games, she ignited debate in each the US and China as an American who selected to compete for China, her mom’s residence nation. However the controversy hardly harm her success; she’s develop into an enormous movie star in China with thousands and thousands of {dollars} in endorsement offers, and she or he gained three medals, together with two golds, in Beijing. She’s solely 22 years previous and is gearing up for much more success, and stardom, in Milan. —SM

Giorgia Meloni. Italy’s first girl prime minister already occupies an odd place on the world stage: She’s a right-wing populist who enjoys a detailed relationship with President Donald Trump and his administration… however she’s additionally, clearly, a European chief who’s needed to grapple with Trump’s assaults on the continent. The Olympics, Italy’s first since 2006, will put her within the world highlight as soon as once more. (That, and a comically Italian scandal whereby a partisan artwork restorer snuck Meloni’s face right into a church fresco.) —CD

Cultural initiatives. Alongside the athletic occasions, Italy is internet hosting a nationwide Cultural Olympiad with exhibitions, theater and music performances, artist installations, festivals, and workshops. Los Angeles ought to be watching carefully, as native officers are involved that the town is manner behind on planning and fundraising for arts programming for the 2028 Video games. Can LA showcase itself as an leisure mecca and a capital of cultural range, because it did when it final hosted the Video games in 1984? The clock is ticking. —Avishay Artsy

Greening the Video games. Bear in mind in the course of the 2024 Summer time Olympics in Paris when organizers have been getting warmth for a way they approached “greening” the competitors? On the time, organizers centered on nudging behavioral selections, akin to defaulting towards climate-friendly meals (learn: no meat) or no AC, which wasn’t in style. So I used to be excited to listen to that this 12 months’s Winter Olympics centered its vitality on the precise building of the Milano Cortina Olympic Village as a substitute. In response to the architects behind the venture, the village employs a modular design made with low-carbon building strategies and prefabricated facade panels. It may be a mannequin for future Video games. —Izzie Ramirez

“The battle of millimeters.” Final month, a dishonest scandal rocked the ski-jumping world when three workers members on the Norwegian group have been suspended for modifying the crotches of the outfits of two jumpers, together with the reigning Olympic champion. In sports activities like ski leaping, small modifications like further stitching can have massive aerodynamic impacts — a part of what one ski leaping fanatic described to The Athletic as “the battle of millimeters.” We’ll see what which means for Norway’s athletes in these Olympics. —Cameron Peters

Olympic muffins. The perfect TikToks from the Olympic Village don’t have anything to do with sport: They’re the clips of globe-trotting, carb-loading, world-class athletes reviewing mid-range cafeteria meals. Chocolate muffins stole the present in Paris. (Can affirm: These muffins are good.) And whereas the Milan Cortina Video games have but to supply a breakout foodstuff, the torta della nonna appears to be like like a contender. —CD

The Minions man. Spanish determine skater Tomàs-Llorenç Guarino Sabaté just isn’t anticipated to medal at this 12 months’s Video games, however he’s already a winner in a single regard: He triumphed over movie manufacturing firm Common Footage in a petty, comical dispute the web dubbed #Minionsgate. Sabaté, who skates to music from the Minions films, was briefly blocked from utilizing his signature soundtrack over copyright considerations, which the Olympics have taken way more severely since a 2022 lawsuit. —CD

AI music on ice. Determine skating is making an attempt to remain related by encouraging determine skaters to make modern music selections: classical music out, pop bangers in. Nevertheless it’s led to a music licensing nightmare within the sport, as Minion Man came upon the onerous manner. Some skaters are turning to an ethically doubtful answer: AI music, which avoids costly charges and extended negotiations. As AI slop hits Spotify playlists and Billboard charts, will issues get sloppy on the ice too? We’ll be watching, er, listening to see. —Peter Balonon-Rosen

New sports activities tech. Gymnastics judges for the 2024 Paris Video games received a leg up from a novel instrument: an AI system that would routinely seize, mannequin, and consider athletes’ actions towards normal parameters. AI instruments can even be in use on the Milan Cortina Video games, the place they’ll do issues like seize the peak, size, and air time of determine skaters’ jumps. Numerous groups have additionally enthusiastically embraced AI-informed coaching… although solely time will inform if that provides them the sting that some have claimed. —CD

NHL stars. Males’s hockey will probably be a very splashy occasion this 12 months: It’s the primary 12 months in additional than a decade that the Nationwide Hockey League has allowed its gamers to take part within the Video games. And following the breakout success of the Canadian homosexual sports activities romance collection Heated Rivalry, there’ll be a brand new viewers of followers and informal viewers that will probably be tuning in. (Thanks, Shane Hollander!) —Christian Paz

The rise of sports activities betting. This 12 months’s Olympics arguably mark the primary occasion since prediction markets like Kalshi and Polymarket turned a mainstream phenomenon. Already, Polymarket customers have traded thousands and thousands of {dollars}’ price of bets on particular person competitions and total medal counts. Notably, the Milan Cortina Video games come simply as these platforms face a wave of new scrutiny associated to sports activities dishonest scandals. —CD

Well being and wellness. It’s more and more onerous to consider that, a mere 5 years in the past, the Tokyo Olympics mandated masking, social distancing, and different Covid protocols. Now, most Olympic athletes don’t even masks routinely throughout journey. Personally, I’d be so confused about sickness. Think about how terrible it might be to get sick whereas sharing a tiny dorm room along with your teammates or not feeling 100% wholesome in the course of the largest second of your profession. Organizers already needed to push the ladies’s hockey sport between Canada and Finland due to a norovirus outbreak. —Lauren Katz

Transferring inductions to early morning may shorten labour by 6 hours

Science

Dr. Mike

February 6, 2026

Transferring inductions to early morning may shorten labour by 6 hours

If given the selection, choosing a labour induction within the early morning may pace issues alongside

Iuliia Burmistrova/Getty Photos

One of the best time to induce labour is within the early morning, analysis suggests. By aligning with our pure physique clocks, early-morning inductions appear to shorten labour occasions and scale back the necessity for Caesarean sections.

“It’s a easy, no-cost method that might make the expertise higher for everybody – the mom, the child and the medical employees,” says Hanne Hoffmann at Michigan State College.

A few third of labours within the UK, the US and Australia at the moment are induced, that means they’re introduced on with medicine or different synthetic means, somewhat than ready for them to begin on their very own. Inductions are generally really useful if a child is overdue or has points with its progress, since immediate supply reduces the danger of stillbirth. Different causes embody the waters breaking with out initiating labour, which may improve the danger of an infection.

The issue is that induced labours are sometimes extra drawn out than people who happen spontaneously. “I’ve had pals who’ve been induced they usually’ve been in labour for 2 days. I used to be simply blown away by how lengthy and painful it was for them,” says Hoffmann.

This led Hoffmann – who research circadian rhythms, pure oscillations within the exercise of our tissues, pushed by inside clocks – to wonder if there is perhaps an optimum time of day for inducing labour. “We all know that spontaneous labour follows circadian patterns as a result of uterine contractions are inclined to peak within the late night, and we primarily give start at night time,” she says. This sample could have advanced as a result of there’s much less risk from predators at night time.

To be taught extra, Hoffmann and her colleagues analysed information of greater than 3000 induced labours performed at a hospital in Michigan between 2019 and 2022. They discovered that the shortest labours have been these induced between 3am and 9am. For instance, these induced at 5am lasted for 15 hours on common, whereas these induced at 11pm lasted a staggering 6 hours longer, about 21 hours on common. The shorter morning-inducted labours have been additionally much less prone to result in emergency Caesarean sections.

The explanation for this early-morning profit could also be that receptors within the uterus are extra conscious of the hormone oxytocin right now. In spontaneous labour, oxytocin is accountable for stimulating uterine contractions. In inductions, an artificial type of oxytocin is normally given to attain the identical impact. “When docs induce labour with an enormous oxytocin bolus, it is perhaps high-fiving the physique’s personal built-in morning oxytocin celebration, giving contractions an additional enthusiastic push,” says Satchidananda Panda on the Salk Institute for Organic Research in San Diego, California.

The researchers additionally discovered that individuals with a excessive physique mass index or who have been giving start for the primary time appeared to profit most from early-morning induction. They now plan to analyze the underlying organic mechanisms.

Whereas it is probably not sensible for hospitals to schedule all inductions between 3am and 9am, they may contemplate prioritising first-time moms and those that are obese or overweight, says Hoffmann.

Importantly, the research didn’t discover any medical issues related to inducing early within the day. “There have been no elevated dangers of admissions to the NICU [neonatal intensive care unit] or different opposed outcomes,” says crew member Rene Cortese on the College of Kansas Medical Middle. “That is one other necessary message of the research: you’re mitigating one threat however not creating one other.”

The crew now hopes to conduct a research to substantiate that assigning pregnant folks to early-morning inductions does certainly enhance their labours. “We have to do a proof of idea research to ensure we will replicate this [finding],” says Hoffmann.

Different chronotherapies – medical interventions which can be timed to align with circadian rhythms – are additionally being explored within the fields of oncology, cardiology and psychiatry. For instance, a latest research discovered that treating most cancers earlier than 3pm may assist sufferers dwell longer.

Matters:

Tech Stack for Vibe Coding Trendy Purposes

Machine Learning

Dr. Mike

February 6, 2026

Tech Stack for Vibe Coding Trendy Purposes

Picture by Creator

I used to hate vibe coding. I believed I may write higher code, design cleaner methods, and make extra considerate architectural choices alone. For a very long time, that was most likely true. Over time, issues modified. AI brokers improved considerably. MCP servers, Claude abilities, agent workflows, planning-first execution, and long-horizon coding instruments turned vibe coding from a gimmick right into a sensible method to construct actual methods.

Sooner or later, I gave in. Since then, I’ve been utilizing Claude Code and OpenCode to construct methods that I might usually contemplate too advanced for speedy iteration. These embody fee platforms, stablecoin wallets, ebook studying functions, and full-stack production-ready net methods. What shocked me was not simply the pace, however the consistency of outcomes as soon as the fitting construction and instruments had been in place.

The actual downside most individuals face with vibe coding shouldn’t be writing code.

It’s selecting the best tech stack.

Frontend, backend, authentication, databases, storage, e mail, funds, and deployment all include numerous choices. With out a clear stack, even sturdy AI coding brokers battle to make good choices. When an agent is given a well-defined and opinionated tech stack, it will probably reliably construct an end-to-end utility with far much less friction.

That’s what this text focuses on.

I’ll stroll via my go-to tech stack for vibe coding trendy functions. That is the stack I reuse as a result of it really works. We’ll cowl how the items match collectively, what every instrument is chargeable for, and the way to go from zero to a production-ready system. I may even share a Claude Code pattern immediate and non-obligatory instruments that assist flip a working prototype right into a system prepared for manufacturing.

# My Tech Stack and Why This Stack Works

On this part, I’ll break down the precise tech stack I exploit for vibe coding and clarify why these instruments work so effectively collectively when constructing actual functions.

The Stack

Subsequent.js (App Router) -> Frontend and layouts
shadcn/ui -> UI parts
Server Actions + Routes -> Backend logic and webhooks
Supabase -> Database, auth, storage
Resend -> Transactional emails
Stripe -> Funds and subscriptions
Vercel -> Deployment and previews

This stack is deliberately easy. You should use the free tier of each platform listed right here, which makes it very best if you’re beginning out. You may deploy a full utility to Vercel free of charge and join managed providers with out upfront value.

Why This Stack Scales With out Getting Messy

Two Subsequent.js primitives do a lot of the heavy lifting because the app grows:

Server Actions deal with type submissions and server-side mutations. They maintain information writes near the UI and take away loads of boilerplate that usually seems early.
Route Handlers deal with webhooks, well being checks, and customized endpoints. This offers you a clear place for exterior methods to speak to your app with out polluting your UI logic.

Supabase provides you database, authentication, and storage with a safety mannequin that lives near the information. When you allow Row Degree Safety early, authorization stays constant because the system grows and also you keep away from painful migrations later.

Resend retains transactional e mail easy and environment-driven.

Stripe Checkout paired with webhooks provides you a dependable method to convert funds into actual entitlements as an alternative of scattered characteristic flags.

Vercel retains preview and manufacturing deployments aligned, so you’re testing in actual environments from day one.

This stack works effectively for vibe coding as a result of it’s opinionated, predictable, and straightforward for an AI coding agent to cause about. As soon as the boundaries are clear, the system virtually builds itself.

# Construct Plan from Zero to a Paid MVP

This construct plan is designed for vibe coding with actual instruments. The objective is to get a production-ready skeleton first, then add functionality in small phases with out breaking earlier choices. Every part maps on to the stack you’re utilizing, so an AI coding agent can comply with it finish to finish.

Section 1: MVP Basis

Construct the total product loop with minimal scope.

Arrange Subsequent.js (App Router) challenge with Vercel deployment
Dashboard shell and navigation utilizing shadcn/ui
Authentication flows utilizing Supabase Auth (signup, login, reset)
One core user-owned desk in Supabase Postgres
CRUD screens powered by Subsequent.js Server Actions
Preview deployments on each change through Vercel

On the finish of this part, you have already got a usable app operating in manufacturing, even when the characteristic set is small.

Section 2: Information Security and Entry Management

Lock down consumer information earlier than including extra options.

Allow Row Degree Safety on user-owned tables in Supabase
Outline learn and write insurance policies primarily based on possession
Use constant patterns like owner_id, created_at, updated_at
Validate entry guidelines via actual UI flows, not simply SQL

This part prevents future rewrites and retains safety aligned with how the app really works.

Section 3: E mail and Storage

Add belief and file dealing with.

Transactional emails through Resend (welcome, verification, resets)
Non-public storage buckets utilizing Supabase Storage
Add flows that respect the identical possession guidelines as your database
Signed URLs or managed entry primarily based on consumer identification

That is the place the product begins to really feel full as an alternative of experimental.

Section 4: Billing and Entitlements

Flip utilization into income.

Create Stripe Checkout periods and redirect customers
Deal with Stripe webhooks with Subsequent.js Route Handlers
Retailer subscription or buy state in Supabase
Gate premium options primarily based on saved entitlements
Make webhook dealing with idempotent utilizing processed occasion IDs

By the tip of this part, you have got a paid MVP that may scale with out altering core structure.

# Claude Code Starter Immediate

You may substitute “Ebook Store + Reader MVP” with your individual thought utilizing the identical Claude Code immediate.

Construct a **Ebook Store + Reader MVP** utilizing this stack:
- Subsequent.js App Router  
- shadcn/ui  
- Supabase (Postgres, Auth, Storage)  
- Resend  
- Stripe (Checkout + webhooks)  
- Vercel  

## Purpose
Ship a production-ready Ebook Store and Reader with paid entry.

## Construct
- Public pages: touchdown, pricing, ebook checklist  
- Auth: enroll, check in, reset password  
- Protected app: reader dashboard  

## Information
- `books`, `chapters`
- Row Degree Safety so customers entry solely their very own information

## Options
- CRUD through Server Actions  
- Reader view with progress monitoring  
- Non-public storage for ebook belongings  
- Welcome e mail  
- Stripe Checkout + webhook-based entitlements  

## Output
- Clear app construction
- Minimal dependencies
- README with setup, env vars, migrations, Stripe, and Vercel steps
- Guide verification guidelines per characteristic

All it’s worthwhile to do is swap Claude Code to Plan Mode, paste the immediate, and alter the thought or regulate the scope primarily based in your wants.

When you begin, Claude will plan the system first after which start constructing step-by-step with out friction. It’ll additionally information you thru organising required providers, creating accounts on third-party platforms, and producing API keys the place wanted.

This makes it straightforward to go from an thought to a working utility with out getting caught on setup or choices.

# Optionally available Instruments

These instruments are usually not required to ship the primary model, however they make it easier to take a look at, monitor, and harden the appliance because it grows in actual utilization.

Class	Instrument choices	What it helps with	When so as to add it
Unit assessments	Vitest	Quick assessments for utilities and server logic	As soon as core CRUD works
Element assessments	React Testing Library	Catch UI regressions in types and states	After the dashboard stabilizes
Finish-to-end assessments	Playwright	Full consumer flows: signup → create → pay	Earlier than including extra options
Error monitoring	Sentry	Stack traces, launch well being, alerting	As quickly as actual customers arrive
Logs	Axiom or Logtail	Searchable request logs, webhook debugging	When webhooks and billing go dwell
Efficiency checks	Lighthouse (CI)	Catch sluggish pages and outsized bundles	Earlier than advertising launches
Schema and migrations	Drizzle Equipment or SQL migrations	Repeatable schema modifications	The second you have got 2+ tables
Background jobs	Inngest or Set off.dev	Async work: emails, exports, cleanup	When workflows develop past requests
Fee limiting	Upstash Redis (or related)	Defend auth endpoints and webhooks	When visitors turns into actual
Product analytics	PostHog (or related)	Funnels, activation, characteristic utilization	After what you measure

# Ultimate Ideas

Trendy growth and engineering instruments are evolving quick. Most of them are actually designed with AI integration in thoughts, providing good documentation, APIs, and MCP-style entry so AI brokers can work with them immediately and construct software program quicker than ever.

If you’re a knowledge scientist who has by no means touched net growth, or a whole newbie who desires to construct one thing actual or launch a startup, I strongly advocate beginning with this tech stack. It requires minimal setup and allows you to deploy a working utility virtually instantly.

It took me practically three months to check and evaluate instruments earlier than selecting this stack. Beginning right here will prevent that point.

If you would like extra flexibility later, you may cut up issues out. For instance, use Neon for the database, Clerk for authentication, and maintain every part else the identical. Spreading tasks throughout instruments makes it simpler to exchange one half with out breaking the remaining as your system grows.

Begin easy, ship early, and evolve solely when it’s worthwhile to.

Abid Ali Awan (@1abidaliawan) is a licensed information scientist skilled who loves constructing machine studying fashions. Presently, he’s specializing in content material creation and writing technical blogs on machine studying and information science applied sciences. Abid holds a Grasp’s diploma in expertise administration and a bachelor’s diploma in telecommunication engineering. His imaginative and prescient is to construct an AI product utilizing a graph neural community for college kids combating psychological sickness.

Posit AI Weblog: Simple PixelCNN with tfprobability

Artificial Intelligence

Dr. Mike

February 6, 2026

Posit AI Weblog: Simple PixelCNN with tfprobability

We’ve seen fairly just a few examples of unsupervised studying (or self-supervised studying, to decide on the extra appropriate however much less
standard time period) on this weblog.

Typically, these concerned Variational Autoencoders (VAEs), whose attraction lies in them permitting to mannequin a latent house of
underlying, impartial (ideally) components that decide the seen options. A doable draw back could be the inferior
high quality of generated samples. Generative Adversarial Networks (GANs) are one other standard method. Conceptually, these are
extremely engaging resulting from their game-theoretic framing. Nonetheless, they are often tough to coach. PixelCNN variants, on the
different hand – we’ll subsume all of them right here below PixelCNN – are usually recognized for his or her good outcomes. They appear to contain
some extra alchemy although. Below these circumstances, what might be extra welcome than a straightforward means of experimenting with
them? By TensorFlow Chance (TFP) and its R wrapper, tfprobability, we now have
such a means.

This publish first provides an introduction to PixelCNN, concentrating on high-level ideas (leaving the main points for the curious
to look them up within the respective papers). We’ll then present an instance of utilizing tfprobability to experiment with the TFP
implementation.

PixelCNN rules

Autoregressivity, or: We’d like (some) order

The fundamental concept in PixelCNN is autoregressivity. Every pixel is modeled as relying on all prior pixels. Formally:

[p(mathbf{x}) = prod_{i}p(x_i|x_0, x_1, …, x_{i-1})]

Now wait a second – what even are prior pixels? Final I noticed one photos had been two-dimensional. So this implies now we have to impose
an order on the pixels. Generally this will likely be raster scan order: row after row, from left to proper. However when coping with
colour photos, there’s one thing else: At every place, we even have three depth values, one for every of purple, inexperienced,
and blue. The unique PixelCNN paper(Oord, Kalchbrenner, and Kavukcuoglu 2016) carried by way of autoregressivity right here as properly, with a pixel’s depth for
purple relying on simply prior pixels, these for inexperienced relying on these similar prior pixels however moreover, the present worth
for purple, and people for blue relying on the prior pixels in addition to the present values for purple and inexperienced.

[p(x_i|mathbf{x}

Here, the variant implemented in TFP, PixelCNN++(Salimans et al. 2017) , introduces a simplification; it factorizes the joint
distribution in a less compute-intensive way.

Technically, then, we know how autoregressivity is realized; intuitively, it may still seem surprising that imposing a raster
scan order “just works” (to me, at least, it is). Maybe this is one of those points where compute power successfully
compensates for lack of an equivalent of a cognitive prior.

Masking, or: Where not to look

Now, PixelCNN ends in “CNN” for a reason – as usual in image processing, convolutional layers (or blocks thereof) are
involved. But – is it not the very nature of a convolution that it computes an average of some sorts, looking, for each
output pixel, not just at the corresponding input but also, at its spatial (or temporal) surroundings? How does that rhyme
with the look-at-just-prior-pixels strategy?

Surprisingly, this problem is easier to solve than it sounds. When applying the convolutional kernel, just multiply with a
mask that zeroes out any “forbidden pixels” – like in this example for a 5×5 kernel, where we’re about to compute the
convolved value for row 3, column 3:

[left[begin{array}
{rrr}
1 & 1 & 1 & 1 & 1
1 & 1 & 1 & 1 & 1
1 & 1 & 1 & 0 & 0
0 & 0 & 0 & 0 & 0
0 & 0 & 0 & 0 & 0
end{array}right]
]

This makes the algorithm sincere, however introduces a distinct downside: With every successive convolutional layer consuming its
predecessor’s output, there’s a repeatedly rising blind spot (so-called in analogy to the blind spot on the retina, however
positioned within the prime proper) of pixels which are by no means seen by the algorithm. Van den Oord et al. (2016)(Oord et al. 2016) repair this
through the use of two completely different convolutional stacks, one continuing from prime to backside, the opposite from left to proper.

Conditioning, or: Present me a kitten

To date, we’ve at all times talked about “producing photos” in a purely generic means. However the actual attraction lies in creating
samples of some specified kind – one of many courses we’ve been coaching on, or orthogonal info fed into the community.
That is the place PixelCNN turns into Conditional PixelCNN(Oord et al. 2016), and it’s also the place that feeling of magic resurfaces.
Once more, as “basic math” it’s not laborious to conceive. Right here, (mathbf{h}) is the extra enter we’re conditioning on:

[p(mathbf{x}| mathbf{h}) = prod_{i}p(x_i|x_0, x_1, …, x_{i-1}, mathbf{h})]

However how does this translate into neural community operations? It’s simply one other matrix multiplication ((V^T mathbf{h})) added
to the convolutional outputs ((W mathbf{x})).

[mathbf{y} = tanh(W_{k,f} mathbf{x} + V^T_{k,f} mathbf{h}) odot sigma(W_{k,g} mathbf{x} + V^T_{k,g} mathbf{h})]

(In the event you’re questioning concerning the second half on the best, after the Hadamard product signal – we gained’t go into particulars, however in a
nutshell, it’s one other modification launched by (Oord et al. 2016), a switch of the “gating” precept from recurrent neural
networks, corresponding to GRUs and LSTMs, to the convolutional setting.)

So we see what goes into the choice of a pixel worth to pattern. However how is that call really made?

Logistic combination chance , or: No pixel is an island

Once more, that is the place the TFP implementation doesn’t observe the unique paper, however the latter PixelCNN++ one. Initially,
pixels had been modeled as discrete values, selected by a softmax over 256 (0-255) doable values. (That this really labored
looks as if one other occasion of deep studying magic. Think about: On this mannequin, 254 is as removed from 255 as it’s from 0.)

In distinction, PixelCNN++ assumes an underlying steady distribution of colour depth, and rounds to the closest integer.
That underlying distribution is a mix of logistic distributions, thus permitting for multimodality:

[nu sim sum_{i} pi_i logistic(mu_i, sigma_i)]

General structure and the PixelCNN distribution

General, PixelCNN++, as described in (Salimans et al. 2017), consists of six blocks. The blocks collectively make up a UNet-like
construction, successively downsizing the enter after which, upsampling once more:

In TFP’s PixelCNN distribution, the variety of blocks is configurable as num_hierarchies, the default being 3.

Every block consists of a customizable variety of layers, known as ResNet layers because of the residual connection (seen on the
proper) complementing the convolutional operations within the horizontal stack:

In TFP, the variety of these layers per block is configurable as num_resnet.

num_resnet and num_hierarchies are the parameters you’re more than likely to experiment with, however there are just a few extra you possibly can
try within the documentation. The variety of logistic
distributions within the combination can be configurable, however from my experiments it’s greatest to maintain that quantity relatively low to keep away from
producing NaNs throughout coaching.

Let’s now see an entire instance.

Finish-to-end instance

Our playground will likely be QuickDraw, a dataset – nonetheless rising –
obtained by asking individuals to attract some object in at most twenty seconds, utilizing the mouse. (To see for your self, simply try
the web site). As of as we speak, there are greater than a fifty million cases, from 345
completely different courses.

Firstly, these knowledge had been chosen to take a break from MNIST and its variants. However similar to these (and plenty of extra!),
QuickDraw could be obtained, in tfdatasets-ready type, through tfds, the R wrapper to
TensorFlow datasets. In distinction to the MNIST “household” although, the “actual samples” are themselves extremely irregular, and infrequently
even lacking important elements. So to anchor judgment, when displaying generated samples we at all times present eight precise drawings
with them.

Getting ready the information

The dataset being gigantic, we instruct tfds to load the primary 500,000 drawings “solely.”

To hurry up coaching additional, we then zoom in on twenty courses. This successfully leaves us with ~ 1,100 – 1,500 drawings per
class.

# bee, bicycle, broccoli, butterfly, cactus, # frog, guitar, lightning, penguin, pizza, # rollerskates, sea turtle, sheep, snowflake, solar, # swan, The Eiffel Tower, tractor, practice, tree courses <- c(26, 29, 43, 49, 50, 125, 134, 172, 218, 225, 246, 255, 258, 271, 295, 296, 308, 320, 322, 323 ) classes_tensor <- tf$forged(courses, tf$int64) train_ds <- train_ds %>% dataset_filter( operate(file) tf$reduce_any(tf$equal(classes_tensor, file$label), -1L) )

The PixelCNN distribution expects values within the vary from 0 to 255 – no normalization required. Preprocessing then consists
of simply casting pixels and labels every to float:

preprocess <- operate(file) { file$picture <- tf$forged(file$picture, tf$float32) file$label <- tf$forged(file$label, tf$float32) checklist(tuple(file$picture, file$label)) } batch_size <- 32 practice <- train_ds %>% dataset_map(preprocess) %>% dataset_shuffle(10000) %>% dataset_batch(batch_size)

Creating the mannequin

We now use tfd_pixel_cnn to outline what would be the
loglikelihood utilized by the mannequin.

dist <- tfd_pixel_cnn( image_shape = c(28, 28, 1), conditional_shape = checklist(), num_resnet = 5, num_hierarchies = 3, num_filters = 128, num_logistic_mix = 5, dropout_p =.5 ) image_input <- layer_input(form = c(28, 28, 1)) label_input <- layer_input(form = checklist()) log_prob <- dist %>% tfd_log_prob(image_input, conditional_input = label_input)

This tradition loglikelihood is added as a loss to the mannequin, after which, the mannequin is compiled with simply an optimizer
specification solely. Throughout coaching, loss first decreased shortly, however enhancements from later epochs had been smaller.

mannequin <- keras_model(inputs = checklist(image_input, label_input), outputs = log_prob) mannequin$add_loss(-tf$reduce_mean(log_prob)) mannequin$compile(optimizer = optimizer_adam(lr = .001)) mannequin %>% match(practice, epochs = 10)

To collectively show actual and faux photos:

for (i in courses) { real_images <- train_ds %>% dataset_filter( operate(file) file$label == tf$forged(i, tf$int64) ) %>% dataset_take(8) %>% dataset_batch(8) it <- as_iterator(real_images) real_images <- iter_next(it) real_images <- real_images$picture %>% as.array() real_images <- real_images[ , , , 1]/255 generated_images <- dist %>% tfd_sample(8, conditional_input = i) generated_images <- generated_images %>% as.array() generated_images <- generated_images[ , , , 1]/255 photos <- abind::abind(real_images, generated_images, alongside = 1) png(paste0("draw_", i, ".png"), width = 8 * 28 * 10, top = 2 * 28 * 10) par(mfrow = c(2, 8), mar = c(0, 0, 0, 0)) photos %>% purrr::array_tree(1) %>% purrr::map(as.raster) %>% purrr::iwalk(plot) dev.off() }

From our twenty courses, right here’s a alternative of six, every exhibiting actual drawings within the prime row, and faux ones beneath.

We most likely wouldn’t confuse the primary and second rows, however then, the precise human drawings exhibit monumental variation, too.
And nobody ever stated PixelCNN was an structure for idea studying. Be at liberty to mess around with different datasets of your
alternative – TFP’s PixelCNN distribution makes it simple.

Wrapping up

On this publish, we had tfprobability / TFP do all of the heavy lifting for us, and so, might deal with the underlying ideas.
Relying in your inclinations, this may be a great state of affairs – you don’t lose sight of the forest for the bushes. On the
different hand: Must you discover that altering the offered parameters doesn’t obtain what you need, you have got a reference
implementation to begin from. So regardless of the consequence, the addition of such higher-level performance to TFP is a win for the
customers. (In the event you’re a TFP developer studying this: Sure, we’d like extra :-)).

To everybody although, thanks for studying!

Oord, Aaron van den, Nal Kalchbrenner, and Koray Kavukcuoglu. 2016. “Pixel Recurrent Neural Networks.” CoRR abs/1601.06759. http://arxiv.org/abs/1601.06759.

Oord, Aaron van den, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. 2016. “Conditional Picture Era with PixelCNN Decoders.” CoRR abs/1606.05328. http://arxiv.org/abs/1606.05328.

Salimans, Tim, Andrej Karpathy, Xi Chen, and Diederik P. Kingma. 2017. “PixelCNN++: A PixelCNN Implementation with Discretized Logistic Combination Probability and Different Modifications.” In ICLR.

Get pleasure from this weblog? Get notified of latest posts by electronic mail:

Posts additionally out there at r-bloggers

45-year-old Chad Michael Watts loses combat with teenage woman at Texas anti-ICE walkout, will get arrested

Technology

Dr. Mike
-

February 6, 2026
0

A forty five-year-old MAGA hat wearer has been arrested after getting out of his truck to <a href="http://<p><robust>Beforehand:</robust><br>• <a href="https://boingboing.internet/2026/01/13/jazz-musicians-disrupt-arizona-republicans-pro-ice-press-event.html">Protesters interrupt Arizona pro-ICE press convention</a><br>• <a href="https://boingboing.internet/2018/03/14/enough-2.html">Scenes from right this moment's nationwide gun management pupil walkout</a><br>• <a href="https://boingboing.internet/2023/02/24/students-across-florida-walk-out-in-protest-of-gov-ron-desantis-anti-american-policies.html">College students throughout Florida stroll out in protest of Gov. Ron DeSantis' anti-American insurance policies</a>

combat teenage women at an anti-ICE protest in Buda, Texas.

Chad Michael Watts was charged with two counts of assault inflicting bodily harm after police decided he was the “major aggressor” in a confrontation with college students from Johnson Excessive College on Monday, experiences KXAN. — Learn the remaining

The put up 45-year-old Chad Michael Watts loses combat with teenage woman at Texas anti-ICE walkout, will get arrested appeared first on Boing Boing.

1...848586...394 Page 85 of 394

Management Pitfall	How Leaders Ought to Deal with It
Leaders who automate duties excessively with out aligning them to organizational targets usually face inefficiencies and wasted funding whereas failing to generate significant enterprise affect.	Align AI initiatives with strategic targets, prioritize high-value workflows, and consider ROI earlier than scaling automation.
Viewing AI solely as a technical venture limits strategic worth as a result of management involvement is essential for driving organization-wide adoption and enterprise alignment.	Make AI a management accountability, contain executives in technique, and guarantee initiatives help organizational targets.
Failing to interact staff or talk advantages can breed resistance and scale back adoption charges which finally undermines the success of AI transformations.	Implement structured change administration, talk advantages clearly, present coaching, and contain groups in AI adoption.
Implementing AI with out strong and well-governed knowledge results in unreliable insights and flawed decision-making together with potential regulatory or moral dangers.	Set up robust knowledge governance, preserve knowledge accuracy and consistency, and monitor AI outputs for bias or errors.
Leaders who don’t actively upskill themselves or their groups threat falling behind evolving applied sciences and failing to extract full worth from AI investments.	Promote steady studying, present AI coaching for leaders and groups, and usually replace abilities to remain forward of expertise.

Why IoT Initiatives Matter for Faculty College students

IoT Undertaking Concepts for Faculty College students

1. Sensible Residence Power Monitoring System

2. Sensible Irrigation System

3. IoT-Primarily based Climate Monitoring System

4. Sensible Parking Administration System

5. Sensible Water High quality Monitoring

6. IoT-Primarily based Attendance System

7. Sensible Avenue Lighting System

8. IoT Well being Monitoring System

9. Sensible Waste Administration System

10. Sensible Hearth Detection System

11. IoT-Primarily based Air High quality Monitoring

12. Sensible Door Lock System

13. IoT-Primarily based Flood Detection System

14. Sensible Fridge Monitoring

15. IoT-Primarily based Automobile Monitoring System

16. Sensible Classroom Automation

17. IoT-Primarily based Sensible Alarm System

18. Sensible Drugs Reminder System

19. IoT-Primarily based Sensible Agriculture Monitoring

20. Sensible Gasoline Leakage Detection System

Learn how to Select the Proper IoT Undertaking

Tutorial Steerage for IoT Initiatives

Conclusion

Resolution overview

Analysis: Immediate enter

Analysis: Immediate design

Analysis: Mannequin selection

Conclusion

Acknowledgements

In regards to the authors

Cloud or autonomy

Why Studying AI is Vital For Leaders in 2026?

Core AI Abilities Leaders Should Grasp in 2026

1. AI Literacy and Strategic Fluency

2. Human–AI Collaboration

3. Knowledge-Pushed Choice Intelligence

4. The Construct–Purchase–Borrow–Bot Expertise Technique

5. Moral Governance & Algorithmic Accountability

6. Adaptive Studying

Roadmap: How Leaders Can Get Began with AI Studying

1. Perceive the Fundamentals of AI and ML

2. Discover AI Use Circumstances Related to Your Business

3. Construct AI-Empowered Choice Abilities

4. Develop Moral and Accountable Management Practices

5. Upskill Groups and Create an AI-Prepared Tradition

Conclusion

# My Tech Stack and Why This Stack Works

The Stack

Why This Stack Scales With out Getting Messy

# Construct Plan from Zero to a Paid MVP

Section 1: MVP Basis

Section 2: Information Security and Entry Management

Section 3: E mail and Storage

Section 4: Billing and Entitlements

# Claude Code Starter Immediate

# Optionally available Instruments

# Ultimate Ideas

PixelCNN rules

Autoregressivity, or: We’d like (some) order

Masking, or: Where not to look

Conditioning, or: Present me a kitten

Logistic combination chance , or: No pixel is an island

General structure and the PixelCNN distribution

Finish-to-end instance

Getting ready the information

Creating the mannequin

Wrapping up