All Courses - Page 10 of 697 - Analytics Campus

Should you nonetheless learn bodily books as an alternative of a Kindle, science says you’ve got these 7 distinct traits

Science

July 23, 2026

Should you nonetheless learn bodily books as an alternative of a Kindle, science says you’ve got these 7 distinct traits

[ad_1]

A printed e-book asks extra of you than a Kindle does. It takes up room, weighs down a bag and can’t summon a brand new title in seconds. But print stays the commonest manner People learn books, at the same time as ebooks and audiobooks have turn into atypical components of the market.

What science has not discovered is a seven-trait persona profile for individuals who desire paper. Studying research often examine comprehension, consideration or reported preferences throughout codecs. They don’t diagnose the character of each particular person holding a paperback.

The seven tendencies under are cautious interpretations of that analysis, not traits assured by the behavior. They describe what a seamless choice for print could mirror, particularly when a reader has easy accessibility to each codecs.

A format choice just isn’t a persona take a look at

The primary complication is that “digital studying” just isn’t one factor. A telephone filled with notifications, a laptop computer used for work and a devoted e-ink Kindle create completely different studying situations. A lot of the analysis compares paper with computer systems or tablets, so its outcomes can’t routinely be utilized to each Kindle reader.

Nor are most individuals loyal to at least one format. A 2026 Pew Analysis Heart survey of US adults discovered print remained extra broadly used than ebooks or audiobooks, however many readers moved between codecs. Value, eyesight, incapacity, journey and library entry can matter greater than persona.

Inside these limits, a number of recurring patterns assist clarify why some readers preserve reaching for paper.

Consideration and reminiscence

1. You defend sustained focus

Print readers usually say that paper helps them focus. In a survey of 10,293 college college students throughout 21 nations, individuals typically most well-liked print for tutorial studying and related it with focus and retention. They most well-liked digital textual content when comfort mattered extra. The examine, printed in PLOS One, relied on self-reported preferences relatively than an goal measure of consideration, and college students studying coursework will not be the identical as adults studying novels for pleasure.

Even so, selecting the much less linked object can operate as a boundary. A paper e-book does one factor. For readers who already know they’re distractible, that constraint could also be a part of the attraction. It suggests a willingness to rearrange the atmosphere round focus relatively than assume consideration will handle itself.

2. You want a bodily map of a narrative

A printed e-book provides landmarks. A reader can do not forget that a scene appeared close to the underside of a left-hand web page, or really feel that solely a skinny stack of pages stays. These cues are much less apparent when each display seems a lot the identical.

In a 2019 experiment, Anne Mangen and colleagues requested individuals to learn the identical lengthy thriller story both in a paperback or on a Kindle. The teams carried out equally on many measures, however the print readers had been higher at reconstructing the chronology of occasions. This was a comparatively small examine of 1 textual content, not proof that paper all the time produces higher reminiscence. It does counsel that readers who favor print could worth the quiet spatial construction a bodily quantity offers.

3. You care whether or not understanding feels actual

Individuals are not all the time good at judging how effectively they’ve realized one thing. In two experiments reported within the Journal of Experimental Psychology: Utilized, Rakefet Ackerman and Morris Goldsmith discovered that individuals studying on screens made much less correct predictions of their efficiency. When examine time was self-regulated, the display group additionally carried out worse than the paper group.

That may be a outcome about studying situations, not an inherent high quality possessed by print lovers. Nonetheless, somebody who repeatedly chooses paper could care concerning the feeling {that a} textual content has been correctly absorbed. The choice can mirror warning about fluency: shifting rapidly by means of phrases just isn’t the identical as understanding them.

Possession and materials attachment

4. You worth possession, not merely entry

A bodily e-book will be lent, annotated, resold, displayed or saved with out an account. An e book is handy, however its use will be formed by a platform’s software program, licensing phrases and system assist. In focus teams concerning the e-book market, researchers discovered that buyers didn’t expertise digital and bodily possession as equal. Restrictions on sharing and management lowered the perceived worth of ebooks for some individuals.

The examine was qualitative, so it can’t inform us how frequent this perspective is. It does reveal a significant distinction. Some print readers will not be rejecting expertise. They’re selecting the model that feels unambiguously theirs.

5. You connect reminiscences and identification to things

Books can mark a interval of life in a manner {that a} filename not often does. A battered novel could carry a prepare ticket, a buddy’s be aware or a espresso stain from the week it was learn. Its worth comes partly from the occasions gathered round it.

Sociologist María Angélica Thumala Olave explored this attachment by means of 43 interviews and 60 written responses in Britain. Contributors described the pleasure of being surrounded by books, the issue of discarding them and using books as objects of reminiscence, identification and connection. This was a qualitative examine of individuals with robust emotions about books, not a consultant persona survey. Its power is in displaying how materials objects can turn into a part of an autobiography.

Why some readers settle for the inconvenience

6. You’ll commerce comfort for a most well-liked expertise

Digital books resolve real issues. They’re searchable, transportable and adjustable, they usually could make studying attainable for individuals who want bigger kind or different accessibility options. Preferring print regardless of these benefits means that effectivity just isn’t the reader’s solely measure of worth.

A meta-analysis by Pablo Delgado and colleagues mixed analysis evaluating comprehension on paper and digital gadgets. It discovered a small common benefit for paper, significantly for informational texts and studying below time stress. The authors didn’t discover that each reader or each textual content benefited. The result’s higher understood as a reminder that friction can generally assist an expertise relatively than spoil it.

A print-preferring reader could also be snug accepting weight, storage and slower acquisition as a result of studying just isn’t merely the switch of knowledge. Texture, place, cowl artwork and the seen progress by means of a quantity are a part of the exercise.

7. You’ve constructed a powerful reader identification

When an individual says “I’m a reader,” the picture they carry could embody cabinets, bookmarks and a e-book within the hand. Format then turns into a part of a practiced identification, not only a technical setting.

A examine of 143 tenth-grade college students discovered that extra avid readers tended to favor paper, whereas reluctant readers had been extra constructive about studying on a tool. An open overview of that examine and associated adolescent analysis concluded that studying expertise helped form format choice. The unique pattern was small, younger and positioned in a number of faculties round Oslo, so the discovering shouldn’t be generalized to all adults. It does, nonetheless, assist the concept that choice grows alongside studying historical past.

A powerful reader identification needn’t imply inflexible loyalty. Somebody could love paper at residence, use a Kindle on a flight and take heed to an audiobook whereas strolling. That flexibility is proof that the underlying dedication is commonly to studying itself.

What the choice actually says

None of those research establishes that individuals who learn print are extra considerate, clever or emotionally deep than individuals who use Kindles. A format selection can’t separate persona from behavior, cash, entry, activity or the easy pleasure of a selected object.

The analysis does assist a narrower conclusion. Bodily books present spatial cues, agency boundaries and a stronger sense of fabric possession. Readers who proceed to decide on them could also be defending focus, valuing management, preserving reminiscences or looking for an expertise that feels completely different from the remainder of life on a display.

Essentially the most revealing trait could due to this fact be neither nostalgia nor resistance to vary. It’s intentionality. These readers have seen the situations below which studying feels richest to them, they usually preserve recreating these situations even when a sooner choice is ready close by.

[ad_2]

Loop Engineering for RAG Technology: Iterate top-k One at a Time

Machine Learning

Dr. Mike

July 23, 2026

Loop Engineering for RAG Technology: Iterate top-k One at a Time

[ad_1]

The person asks “what’s the efficient date of this coverage?”. Retrieval returns 5 candidate line-windows and passes all of them to the LLM in a single immediate. The LLM reads all 5 to extract the identical date the primary candidate already had. Chunks two to 5 have been signatures, footnotes, and a paragraph about historic dates. Paid for nothing. Ship the top-1 first, ask the LLM if that’s sufficient, and cease when it says sure: sequential feeding cuts the token price by 80 % on this class of query. The remainder of the article catalogues the place sequential wins, the place a single-call over all Ok is the best default, and the way the query parser dispatches between the 2.

This text is a companion to Enterprise Doc Intelligence, the collection whose philosophy is specified by Amplify the Professional, the collection that builds enterprise RAG from 4 bricks (doc parsing, query parsing, retrieval, era). It sits between Article 8 (era) and Article 9 (upgrading the mini-RAG) and develops one particular choice: how do you feed the top-Ok retrieved candidates into the era brick?. The reply most pipelines have is “all Ok directly”. This text catalogues the second regime (sequential, top-1 first) and reveals when each wins.

The naive baseline this text pushes again on

The 2 regimes for feeding the top-Ok to the era brick – Picture by writer

Naive RAG ships batch by default. Retrieval returns top-5, the LLM will get all 5, the reply comes again. It really works, and on arduous questions (comparability, itemizing) it’s the proper alternative. However the silent price is paid on each different query: the simple factual ones the place the top-1 already had the reply. This text walks the second regime and the dispatch that picks per query.

*The place this text sits within the collection: brick 8 (era) highlighted – Picture by writer*

📓 The runnable pocket book for this text is on GitHub: doc-intel/notebooks-vol1. It sends the identical questions by means of each batch and sequential era, prints the per-question token price and the 2 sufficiency booleans (answer_found, complete_answer_found) that cease the loop, and reproduces the dispatch desk by yourself machine.

*The general public companion-code repo at doc-intel/notebooks-vol1 – Picture by writer*

1. Two regimes for feeding the top-Ok to era

Retrieval arms era an ordered top-Ok. There are two methods to feed it in, they usually price very in another way.

1.1 The mounted top-Ok pipeline and what it wastes

The default sample throughout most RAG tutorials appears like this:

top_k = retrieval(query, ok=5)
reply = era(query, top_k)

It runs in two steps for each query. Token price is roughly generation_cost(query + 5 chunks of context). Latency is roughly retrieval + one LLM name. And the LLM fortunately processes the 5 chunks even when the top-1 was already adequate and chunks 2..5 added zero data.

Concretely: the person asks “what’s the efficient date of this coverage?”. Retrieval returns 5 line-windows the place the key phrase "efficient" seems. The primary one is the reply (“efficient from January 1, 2026”). Chunks 2..5 are signatures, footnotes, and a paragraph about historic efficient dates of previous insurance policies. The LLM reads all 5 to extract the identical date the primary one already had. On a corpus of 1 doc, the associated fee is rounding error. On a corpus of 50k insurance policies, that is actual cash monthly.

1.2 Sequential: top-1 first, sufficiency predicate, escalate if wanted

The sequential regime treats the Ok candidates as an ordered checklist and asks the era brick to validate sufficiency at every step:

for i, candidate in enumerate(top_k):
    reply = era(query, [candidate])
    if reply.answer_found and reply.complete_answer_found:
        break

The sufficiency sign lives contained in the typed contract Article 8A launched. The AnswerWithEvidence schema exposes answer_found (was the query’s data current on this candidate?) and complete_answer_found (was your complete reply current, not a fraction?). The loop reads these fields, not a customized heuristic.

Within the efficient date instance above: era runs as soon as on the top-1 candidate, the LLM returns answer_found = True, complete_answer_found = True, the loop exits. Tokens spent: generation_cost(query + 1 chunk of context). That’s 1/5 of the batch price for this query, on a pipeline that handles hundreds of comparable lookups per day.

1.3 Batch: ship all Ok directly, let the LLM arbitrate

Batch mode retains the default behaviour: one LLM name sees all Ok candidates and produces one typed reply. Its case is constructed on three query sorts the place sequential breaks down:

Price evaluation: batch at all times pays the generation_cost(query + Ok chunks of context) as soon as. Sequential pays generation_cost(query + 1 chunk of context) within the simple case (top-1 adequate) and generation_cost(query + 1 chunk) × Ok within the worst case (each candidate inadequate). On a typical enterprise corpus with Ok = 5, sequential is cheaper on common for factual lookups (~80% of typical site visitors) and costlier for itemizing / comparability (~20%).

2. The dispatch choice: per query, not per pipeline

The clear structure doesn’t decide batch or sequential globally. It picks per query, utilizing the parsed question_df row from brick 2. Query form, decomposition sample, and intent drive the selection:

*4 query shapes, 4 routing choices; the parser fills two columns, the dispatcher reads them – Picture by writer*

The dispatcher reads question_df.answer_shape and question_df.decomposition and routes. Naive RAG has no method to make this distinction as a result of it has no parsed query to learn from.

The identical routing desk holds throughout sectors and professions. Totally different domains carry the identical form patterns and the identical sequential / batch choice flows out of them:

*The dispatch logic applies whether or not the corpus is insurance coverage, authorized, medical, monetary, or compliance – Picture by writer*

In each row, the sequential column is a single typed worth (Quantity, Date, Boolean) and advantages from the top-1 cease. The batch column is a listing or a comparability and wishes all Ok candidates seen directly. The dispatcher’s desk covers all 5 sectors with the identical logic.

3. The sufficiency sign

Sequential mode leans on one factor: the era brick reporting whether or not the candidate it simply learn was sufficient. That sign, and the foundations that cease the loop, stay right here.

3.1 The place the sufficiency sign lives within the typed contract

Sequential mode solely works if the era brick can self-report whether or not the candidate it simply noticed contained the reply. The typed contract from Article 8A is what makes this doable:

class AnswerWithEvidence(BaseModel):
    worth: Any
    proof: checklist[Span]
    answer_found: bool
    complete_answer_found: bool
    confidence: float = Subject(ge=0, le=1)
    caveats: checklist[str] = []

The sequential loop reads answer_found and complete_answer_found, not a confidence float or a customized heuristic. The clear separation between the 2 booleans (from Sample 4 of Article 8ter) is what makes the loop deterministic. Discovered + incomplete says “proceed to the following candidate”; discovered + full says “cease”; not discovered says “proceed or quit at Ok”.

A confidence float would power a threshold (e.g. “cease at 0.8”), and that threshold drifts mannequin to mannequin. Two booleans don’t drift.

*Two booleans drive three exits; the center exit loops again to era with the following candidate – Picture by writer*

3.2 Bounded iteration: even sequential has to cease

The sequential loop has three exits, not one:

Naive sequential implementations skip the third exit and burn tokens endlessly on edge circumstances. The collection model units a finances upfront and the dispatcher logs it on each name.

4. Price, and the place the collection stops

Two regimes, two price profiles, and one boundary the collection doesn’t cross.

4.1 A concrete price comparability on a hundred-question batch

To make the trade-off actual, here’s a back-of-envelope for an enterprise insurance coverage Q&A workload:

The ratio is 65% saving on enter tokens for era on this workload. The precise ratio is determined by the easy-question proportion and the chunk measurement; the precept is that sequential is a budget default as soon as the typed contract is in place. Batch is reserved for the query sorts that want it.

4.2 The agentic temptation, and the place the collection stops

A pure subsequent step is to let the LLM determine between batch and sequential per query (agentic dispatch). The collection stops in need of this. The dispatcher in Part 2 is deterministic-dispatcher (one of many three approaches catalogued in Article 6C), not LLM-decides. The reason being the identical one which holds throughout the collection: audit. The identical query on the identical day should route the identical means; an LLM that re-plans the dispatch per name can not give that assure.

In case you want a stronger agentic loop (the LLM picks which candidates to have a look at, in what order, with what scope), the larger article on adaptive RAG loops is the best place. This text retains the scope to the deterministic sequential vs batch choice pushed by the parsed query.

5. The choice belongs to the parser, not the LLM

The mounted top-Ok + batch pipeline is the best default for the query sorts the place each candidate issues. It’s the incorrect default for the factual lookups that make up most enterprise site visitors. The collection provides two items: the typed sufficiency sign (answer_found, complete_answer_found) from Article 8A, and the dispatch desk from Part 2. Collectively they let the era brick cease after the primary candidate when that’s sufficient, and course of all Ok when the query kind requires it. The choice is made by the parser, not by the LLM, which retains the audit path intact.

6. Additional studying and sources

The sequential / batch choice sits on the boundary between retrieval and era. The articles that body both sides:

[ad_2]

Figuring out the ROI of AI requires information that almost all firms lack

Dr. Mike

July 23, 2026

Figuring out the ROI of AI requires information that almost all firms lack

[ad_1]

Making use of classes discovered from managing cloud spend gained’t be a repair for the AI and ROI quandary. True, cloud taught a technology of CFOs that billing with out enterprise context is noise. So to get cloud ROI, they stitched two information sources collectively: price information plus enterprise information. AWS reveals which account, which area, which tag, which useful resource. Merge in buyer and product mappings on high and the ROI of the cloud spend comes into focus.

However AI is tougher. It requires three information sources: price, enterprise, and telemetry—the automated assortment of knowledge from disparate sources that helps to make clear the entire image of what occurred and why. An government or engineering lead can have AI invoices and buyer income. However they don’t have any method to join them to enterprise worth. The token rely on the OpenAI bill doesn’t specify which buyer triggered which name, which function it served, or whether or not the immediate produced a enterprise end result. That information doesn’t exist within the supplier’s billing.

AI suppliers gained’t repair this downside

The state of affairs will not be prone to change anytime quickly as a result of AI suppliers are usually not within the enterprise of attributing an enterprise’s prices to that enterprise’s prospects. As an alternative, AI suppliers are within the enterprise of promoting tokens. The granularity they expose is the granularity their billing techniques require, not the granularity a CFO requires.

[ad_2]

Meet Gigatoken: A Rust BPE Tokenizer that Encodes Textual content at 24.53 GB/s, as much as 989x Quicker than HuggingFace Tokenizers

Artificial Intelligence

Dr. Mike

July 23, 2026

Meet Gigatoken: A Rust BPE Tokenizer that Encodes Textual content at 24.53 GB/s, as much as 989x Quicker than HuggingFace Tokenizers

[ad_1]

Tokenization is the one a part of the language modeling stack that just about no person profiles. Gigatoken, launched by Marcel Rød (a PhD scholar from Stanford) below an MIT license, argues that this was a mistake. The library encodes textual content at gigabytes per second on a single machine, in opposition to baselines which are already multithreaded Rust.

The GPT-2 tokenizer benchmarking yields exceptional outcomes: evaluated on the 11.9 GB owt_train.txt corpus utilizing a 144-core AMD EPYC 9565 dual-socket setup, Gigatoken processes information at a staggering 24.53 GB/s. Compared, OpenAI’s tiktoken achieves 36.0 MB/s, whereas HuggingFace tokenizers registers at 24.8 MB/s on the similar {hardware} configuration. These marks display efficiency benefits of 681x and 989x, respectively.

On an Apple M4 Max with 16 cores, the identical GPT-2 workload runs at 8.79 GB/s, or 1,268x HuggingFace tokenizers and 140x tiktoken. On a client AMD Ryzen 7 9800X3D, it runs at 6.27 GB/s, or 106x and 68x. The speedup shouldn’t be an artifact of 1 CPU or one vocabulary.

What’s Gigatoken

Gigatoken is a byte-pair encoding (BPE) tokenizer written in Rust with Python bindings. It ships on PyPI as gigatoken (model 0.9.0, launched 21 July 2026) and installs with pip set up gigatoken. The repository is 66.2% Rust and 33.3% Python. It helps 23 distinct tokenizer households within the revealed benchmarks, overlaying GPT-2, GPT-OSS, Llama 3 by means of 4, Qwen 2 by means of 3.6, DeepSeek V3/R1/V4, GLM 4 and 5, Kimi K2, Nemotron 3, Phi-4, OLMo 2/3, ModernBERT, Gemma and Mistral.

There are two methods to make use of it. Compatibility mode wraps an current HuggingFace or tiktoken tokenizer and preserves precise output parity, at an actual price to throughput. The creator (Marcel) states on Hacker Information that compatibility mode delivers roughly 200–300x relying on utilization, as a result of it nonetheless pays Python overhead for checklist creation and string-to-bytes conversion. The native Gigatoken API lets Rust learn information instantly and is the place the revealed numbers come from.

The interactive benchmark explorer

Each quantity under is drawn from the repository’s benchmarks part and the pretokenizer optimization log. Swap CPUs, stroll the optimization historical past, or estimate how lengthy your personal corpus would take.

‘;
}
operate sprint(){ return ‘—‘; }

operate renderBench(key)>. ‘
+ ‘Dimmed bars are SentencePiece-based tokenizers, that are much less optimized in gigatoken. ‘
+ ‘tiktoken rows are crammed solely the place official assist exists.

‘;
return h;

operate renderPerf(){
var max = 1049, h=”

”
+ ‘Single-threaded GPT-2 pretokenizer throughput on 100 MB of OpenWebText, Apple Silicon, cargo bench with lto = “fats”. ‘
+ ‘A flowery-regex baseline runs the identical pretokenizer at ~47 MiB/s.

‘;
PERF.forEach(operate(s){
h += ‘

‘+s[0]+’

‘
+ ‘

‘+s[1]
+ (s[3]===’dangerous’?’not stored‘:’stored‘)+’‘
+ ‘‘+s[2].toLocaleString()+’ MiB/s

‘
+ bar(s[2],max,s[3]===’dangerous’)
+ ‘

‘+s[4]+’

‘;
});
h += ‘

Internet consequence: 2.27× over the winnow + NEON baseline, 22.3× over the regex implementation.

‘;
return h;
}

operate fmt(sec){
if(sec < 1) return (sec*1000).toFixed(0)+’ ms’;
if(sec < 90) return sec.toFixed(1)+’ s’;
if(sec < 5400) return (sec/60).toFixed(1)+’ min’;
if(sec < 172800) return (sec/3600).toFixed(1)+’ hours’;
return (sec/86400).toFixed(1)+’ days’;
}

operate renderCalc(){
var opts=””;
Object.keys(DATA).forEach(operate(okay){
opts += ‘‘;
});
return ‘

‘
+ ‘

CPU

‘
+ ‘

Tokenizer

‘
+ ‘

Linear extrapolation from the measured throughput above. It assumes the reported price holds ‘
+ ‘for the entire corpus, single machine, no I/O ceiling. Actual pipelines are often storage-bound earlier than they’re tokenizer-bound. ‘
+ ‘For scale: the gigatoken README notes that at EPYC charges you might tokenize all of Frequent Crawl — ‘
+ ‘roughly 130 trillion tokens — in slightly below 6.5 hours.

‘;
}

operate calcUpdate(){
var cpu = R.querySelector(‘#gt-cpu’).worth;
var idx = parseInt(R.querySelector(‘#gt-tok’).worth,10);
var gb = parseInt(R.querySelector(‘#gt-sz’).worth,10);
R.querySelector(‘#gt-szv’).textContent = gb.toLocaleString();
var row = DATA[cpu].rows[idx];
var tg = gb / row[1];
var h=”

gigatoken

“+fmt(tg)
+ ‘

at ‘+row[1].toFixed(2)+’ GB/s

‘;
if(row[2] !== null){
var th = (gb*1000) / row[2];
h += ‘

HF tokenizers

‘+fmt(th)
+ ‘

at ‘+row[2].toFixed(1)+’ MB/s

‘;
h += ‘

Time saved

‘+fmt(th-tg)
+ ‘

‘+row[4]+’× speedup

‘;
} else {
h += ‘

HF tokenizers

—

‘
+ ‘

not reported for this tokenizer

‘;
}
if(row[3] !== null){
var tt = (gb*1000) / row[3];
h += ‘

tiktoken

‘+fmt(tt)
+ ‘

at ‘+row[3].toFixed(1)+’ MB/s

‘;
}
R.querySelector(‘#gt-res’).innerHTML = h;
}

operate fillTok(){
var cpu = R.querySelector(‘#gt-cpu’).worth, sel = R.querySelector(‘#gt-tok’), h=””;
DATA[cpu].rows.forEach(operate(r,i){ h += ‘‘; });
sel.innerHTML = h;
}

var USAGE =
‘

Set up

‘
+ ‘

pip set up gigatoken

‘
+ ‘

Compatibility mode — drop-in, roughly 200–300× per the creator

‘
+ ‘

import gigatoken as gtnn'
  + '# Minimal change from current HuggingFace tokenizers utilizationn'
  + 'hf_tokenizer = ...ntokenizer = gt.Tokenizer(hf_tokenizer).as_hf()nn'
  + 'tokens = tokenizer.encode_batch(["This is a test string", "And here is another"])nn'
  + '# OR with tiktokenntiktokenizer = ...ntokenizer = gt.Tokenizer(tiktokenizer).as_tiktoken()

‘
+ ‘

Gigatoken API — quickest path, Rust reads the information instantly

‘
+ ‘

import gigatoken as gtnn'
  + 'tokenizer = gt.Tokenizer("Qwen/Qwen3-8B")  # accepts HF mannequin namesn'
  + 'file_source = gt.TextFileSource(["owt_train.txt"], separator=b"<|endoftext|>")n'
  + 'tokens = tokenizer.encode_files(file_source)

‘
+ ‘

Verify your personal tokenizer with out putting in something

‘
+ ‘

uvx --with tokenizers gigatoken bench 'openai-community/gpt2' owt_train.txt n'
  + '    --validate --doc-separator "<|endoftext|>"

‘
+ ‘

The benchmark subcommand validates output in opposition to HuggingFace tokenizers and prints a timing comparability. ‘
+ ‘On macOS, run it twice: the primary run triggers a safety scan that slows the Rust code.

‘;

var LIMITS =
‘

Identified points, acknowledged by the creator

WordPiece shouldn’t be supported but. BERT-family WordPiece fashions are out of scope for now.

‘
+ ‘

SentencePiece is barely partly optimized. Gemma, Mistral, CodeLlama and Llama 2 vocabularies land within the 7–22× band, not the 1,000× band.

‘
+ ‘

File sinks should not applied within the Gigatoken API.

‘
+ ‘

Python iteration makes use of ABI3, which is slower than version-specific CPython APIs. Early experiments present a 2× achieve for overhead-bound instances.

‘
+ ‘

Home windows is barely examined. The creator recommends WSL for now.

‘
+ ‘

Compatibility mode prices efficiency. Actual output parity with HuggingFace tokenizers is preserved, however not the complete speedup.

‘
+ ‘

Acknowledged AI use

The creator states many of the code was written by hand, and that AI assisted with the user-facing API, compatibility widening, ‘
+ ‘porting SIMD methods between AVX512 / AVX2 / NEON, the ultimate profiling levels and roughly the final 4× of efficiency, and refactoring.

‘
+ ‘

License and packaging

MIT licensed. Rust 66.2% / Python 33.3% by repository language share. Printed to PyPI as gigatoken.

‘
+ ‘

‘;

operate present(v){
var c = CMDS[v]; if(!c) return;
LINE.textContent = c.line;
BODY.innerHTML = c.get();
R.querySelectorAll(‘.gt-cmd’).forEach(operate(b){
b.setAttribute(‘aria-selected’, b.dataset.v === v ? ‘true’ : ‘false’);
});
if(v === ‘calc’){
fillTok(); calcUpdate();
R.querySelector(‘#gt-cpu’).addEventListener(‘change’, operate(){ fillTok(); calcUpdate(); });
R.querySelector(‘#gt-tok’).addEventListener(‘change’, calcUpdate);
R.querySelector(‘#gt-sz’).addEventListener(‘enter’, calcUpdate);
}
resize();
}

R.querySelectorAll(‘.gt-cmd’).forEach(operate(b){
b.addEventListener(‘click on’, operate(){ present(b.dataset.v); });
});

operate resize(){
attempt{
if(window.dad or mum !== window){
window.dad or mum.postMessage({gigatokenHeight: R.offsetHeight + 40}, ‘*’);
}
}catch(e){}
}
window.addEventListener(‘resize’, resize);

present(‘epyc’);
})();

[ad_2]

Your subsequent WHOOP may very well be free, courtesy of this fancy platinum bank card

Technology

Dr. Mike

July 23, 2026

Your subsequent WHOOP may very well be free, courtesy of this fancy platinum bank card

[ad_1]

Kaitlyn Cimino / Android Authority

TL;DR

Robinhood now features a yearly WHOOP Peak membership as a bank card perk.
WHOOP customers shall be eligible for a $239 credit score for an annual membership, {hardware} included.
The supply is barely legitimate for those who make the acquisition by means of your Robinhood Platinum Visa card.

Regardless of newer options, such because the Fitbit Air, cropping up not too long ago, WHOOP’s screenless bands nonetheless take the highest spot for individuals who care about health metrics greater than something. The companion app, nevertheless, is considerably of a sticking level, not due to what it presents, however the subscription wall WHOOP places as much as allow you to entry your well being knowledge. That’s why some folks have additionally tried to interrupt by means of the wall with an unofficial open-source app for his or her WHOOP bands.

However for those who’d moderately take issues into your personal palms and nonetheless don’t need to pay for a subscription, there’s one other strategy to get a yearly WHOOP subscription with out paying for it. Fintech platform Robinhood not too long ago introduced that individuals who personal its Platinum Visa bank card will now get a complimentary annual WHOOP Peak subscription. Whereas the $239 yearly subscription is not going to be routinely obtainable as a perk with the cardboard, Robinhood clarifies that the subscription price shall be offset as a “assertion credit score.” Which means you could buy your subscriptions by means of a Robinhood card and never different sources. The $239 invoice additionally contains your {hardware} prices, so that you successfully pay nothing.

Based on Robinhood’s up to date phrases of service, you could activate the subscription solely through WHOOP; doing so through a third-party service provider or bundled with any present card makes you ineligible. The credit take eight weeks to be mirrored in your assertion, which suggests you may need to bear the prices for the primary month.

One massive caveat to you getting the advantages is that you could activate the supply from Robinhood’s app earlier than making a purchase order on WHOOP. One other side to notice is that Robinhood’s Platinum card isn’t free itself. You pay a $695 yearly membership charge with the cardboard. The platform then provides you credit, together with $500 to spend on flights and resorts and one other $500 for DoorDash or in-person eating (with month-to-month restrictions). As well as, you stand up to five% cashback on flight and resort bookings, together with different wellness advantages, together with free memberships to Perform Well being, Amazon One Medical, and as much as $1000 in credit for Eight Sleep.

There’s no readability on whether or not this supply is restricted to new subscribers or additionally extends to current ones. If you happen to’re within the latter class, you’ll want to change funds over to your Robinhood card. The supply applies to any purchases made by means of the top of 2027, which suggests the subsequent two years of your WHOOP membership are coated for those who pay upfront. Nonetheless, there’s no assure in regards to the future.

Thanks for being a part of our neighborhood. Learn our Remark Coverage earlier than posting.

[ad_2]

Orcas smash lifeless sunfish into bite-sized items

Science

Dr. Mike

July 23, 2026

Orcas smash lifeless sunfish into bite-sized items

[ad_1]

Regardless of the bloody image painted within the film Jaws, the true king of the ocean just isn’t the nice white shark (Carcharodon carcharias). It’s the orca whale (Orcinus orca). These expert hunters stalk their prey in a wide range of methods, from hydroplaning onto the sand to seize seals to corralling fish into massive bait balls for a snack.

A gaggle of orcas within the Gulf of California seem to have an extra software of their searching toolbox—and it was caught on digital camera. One whale holds down the lifeless physique of a giant sunfish, which may develop over 6 ft lengthy and weigh over 4,000 kilos. The whale then lets go, earlier than one other orca rams into the sunfish carcass at a excessive pace. The collision breaks aside the fish’s tissue into a whole lot of tiny items.

Orca hold-to-ram conduct

“At first, I used to be simply excited as a result of I might see what they had been consuming and doc it,” Dr. Katy Ayres, a analysis scientist from Beneath The Waves and Orcas Mexico and information with Latitude Adventures who captured the footage, tells Well-liked Science. “Typically it’s laborious to determine the prey. However then I wasn’t anticipating the sunfish to fragment into so many items. Because it occurred, it grew to become clear we had been witnessing one thing particular.”

The occasion is detailed in a examine co-authored by Ayres and printed at this time within the journal Frontiers in Ethology.

Items of sunfish in every single place

Sunfish are among the many planet’s heaviest bony fish species. They transfer by means of a number of deeper components of the oceans, however come to the floor to relaxation, regulate their physique temperatures, or discover organisms that may eat the parasites off their pores and skin. Whereas these floor waters will be nice for exfoliation, it makes them extra weak to predators. Within the Gulf of California, they’re generally hunted and eaten by orcas.

“Usually, the orcas right here have an extremely various eating regimen. Whereas some teams seem to specialize on sure prey, as a complete, they’ve been documented searching all the pieces from whales and dolphins to sharks, rays, sea turtles, and naturally the most important bony fishes on the planet, the sunfishes,” says Ayres. “They’re nonetheless a comparatively understudied orca inhabitants, so each new commentary helps fill in one other piece of the puzzle.”

These new observations are primarily based on two impartial searching occasions that occurred about one 12 months aside. In July 2024, Ayres noticed how a gaggle of orcas, together with one juvenile, interacted with a lifeless sharp-tail sunfish (Masturus lanceolatus) that the group had beforehand killed. A feminine orca held the sunfish by its tail fin whereas a male orca accelerated and swam towards each.The feminine launched it shortly earlier than the male rammed into it. The juvenile then ate the smaller bits of sunfish that had exploded as a result of impression, whereas the adults consumed what remained of the physique, however didn’t devour the smaller sunfish fragments.

An analogous occasion in September 2025 was filmed by Héctor Franz and had that very same hold-to-ram approach and feeding conduct.

A household meal

Ramming like this has not been beforehand described as a predation technique by orcas Because the sunfish had been already deceased in each occasions, the ramming could also be a processing approach as a substitute of a approach to ship a killing blow.

“Essentially the most shocking half was how coordinated they had been. One orca held the sunfish in precisely the fitting place and let go on the excellent second for the opposite to strike it with unimaginable power,” says Ayres.

The gelatinous layer that covers sunfish known as a capsule, and their pores and skin hosts all types of distinct microorganisms. If this tissue is shattered, the entire microbes within it unfold into the water and alter the nutrient composition. The workforce believes that the tissue fragmentation might facilitate contact between sunfish and orca microbiomes. These assembly of the microbiomes might give the orca dietary or different advantages together with immune regulation.

The ramming conduct can also be a type of parental funding. Dividing the carcass into extra manageable items for the youthful whales aligns with beforehand noticed behaviors.

“It was additionally actually beautiful to see the younger orca profiting from the smaller items, which raises attention-grabbing questions on whether or not the adults had been making the prey simpler for it to eat,” Ayres remembers.

Enjoying with their meals

Whereas this could possibly be a approach to enhance immunity and assist feed their younger, ramming merely could also be a enjoyable recreation for these extremely social whales.

For the sunfish, the interplay might present new insights into their anatomy and different bodily traits. Thus far, fragmentation of tissue has solely been documented in sharp-tail sunfish. This implies that tissue disintegration could also be a species-specific structural response to high-energy impacts. The tissue in bigger sunfish species could not essentially disintegrate this manner, however we’ll solely know that from future examine.

It’s additionally unclear if the identical orcas had been concerned in each occasions. Extra high-quality footage that captures the orca’s dorsal fins and signature eye patches will help determine people.

“Seeing orcas here’s a actual privilege and a comparatively uncommon expertise. Though there are actually rules, we’re seeing a rise in boats looking for them within the hope of getting individuals within the water and typically forcing encounters,” Ayres concludes. “I feel it’s necessary that guests include lifelike expectations and respect these animals of their pure setting and never simply as a possibility for a photograph or social media put up.”

[ad_2]

The best way to scale agentic AI adoption: A 4-stage studying mannequin

Dr. Mike

July 23, 2026

The best way to scale agentic AI adoption: A 4-stage studying mannequin

[ad_1]

Enterprise corporations don’t develop into agentic by turning on a instrument. They develop into agentic when individuals study, in levels, learn how to direct AI methods towards enterprise motion.

That’s the reason adoption of agentic AI ought to be handled as a studying pathway, a curriculum. As customers construct confidence and functionality, they transfer from asking AI for solutions to orchestrating AI and brokers throughout workflows, with acceptable controls, governance and suggestions.

A helpful means to have a look at it’s akin to a school curriculum of understanding 101, 201, 301 and 401, the place every degree develops the conduct and working self-discipline required for the subsequent. This isn’t a collection of lectures however hands-on analysis, experimentation and tasks so customers come away with the data and expertise that finest match their wants.

Adopting agentic operations by 4 programs

First, by the point most IT leaders actively pursue agentic AI know-how, most individuals within the firm have already had a style of AI as a copilot or assistant. They’re asking AI instruments questions and utilizing them to write down copy and create pictures. The following step goes from merely prompting AI for assist to empowering it to do the work.

Associated:How ecosystem partnerships speed up enterprise AI scale

This leap is not fast. Agentic AI adoption is not a swap to flip; reasonably, it is a curriculum, a development from 101 to 401, the place every degree builds on the final. Embracing the development eases an organization’s tradition into agentic workflows. Extra importantly, it trains and motivates customers to repeatedly use the know-how to succeed in outcomes.

Agentic 101: Asking for motion

The entry-level course on agentic AI begins with the fundamentals, beginning with what most customers already perceive about how they work together with giant language fashions (LLMs): The best way to write prompts that ship outputs.

There is a distinction, although, in how a consumer prompts native agentic AI methods. The questions want to steer brokers towards extra particular actions reasonably than broad search queries. It is useful to consider an AI LLM because the mind and agentic AI because the fingers and toes. Customers have to ask the tech to behave.

In Agentic 101, customers develop comfy interacting with AI brokers. Additionally they study that with out structured information, workflows and permissions, agentic AI cannot reliably execute duties.

Agentic 201: Guiding outcomes

The following degree of studying and expertise is about bettering information entry, strengthening safety and creating trusted environments the place brokers can function and customers can be ok with outcomes. AI brokers require real-time, ruled and enriched information to be fed to them always. Additionally they have to function below strict permissions and tips.

Associated:My AI governance framework is a blast radius: Recommendation from a CEO

Practitioners learn to information AI brokers by setting controls, so AI brokers do not entry sure information units. They pull levers to redirect AI brokers towards information and insights, thus altering how the AI works towards a particular enterprise end result. For instance, a shopper items model could wish to goal a sure area or set of shops, and the consumer can information AI brokers to make selections on pricing, promotions and stock to satisfy their wants.

However customers should be hands-on in collaborating with brokers to get the outcomes they want.

Agentic 301: Constructing repeatable recipes

The 300 degree is the place customers develop into true companions with AI brokers. Practitioners develop into superior sufficient to create profitable repeatable processes, additionally known as expertise or recipes.

Customers develop structured sequences that empower AI brokers to investigate information, apply logic, generate outputs and collaborate with different brokers and methods. Plus, the method is powerful sufficient to be reused, shared and regularly refined.

At a retail model, a advertising workforce would possibly create an “agentic recipe” that standardizes how a promotional marketing campaign is analyzed or how buyer lifetime worth is scored, for instance. It is at this degree the place customers depend on a number of brokers throughout instruments and methods, they usually shift from manually powering AI brokers to orchestrating outcomes.

Associated:The Week of July 6–10: What occurred, what issues, what’s subsequent

Agentic 401: Attaining orchestration

The final part permits a corporation to maneuver into full agentic transformation. Customers who’re comfy defining agentic subtasks can handle and oversee multi-agent orchestration, collaborating inside closed suggestions loops and even reimagining revolutionary methods of working with AI. For instance, an attire retailer may use agentic AI to construct management towers and autonomous brokers to remain forward of provide chain points.

Right here, customers enable brokers to suggest actions, floor insights and suggest new workflows. It is a system of AI brokers not solely performing requested duties but in addition contemplating methods to enhance how duties are approached and executed.

Governance, fixed monitoring of outcomes and human oversight stay key. It is a commencement to a mannequin the place customers work hand in hand with agentic AI.

Curriculum method to information agentic adoption

A curriculum-based method helps organizations and groups construct lasting agentic functionality, not one-time experimentation. It builds a shared language for a way groups work with AI brokers, from primary activity execution to enterprise-scale orchestration.

Agentic success ought to be monitored and measured by outcomes based mostly on how constantly groups and brokers clear up actual enterprise issues collectively. It’s not about who automates quickest; it is about who learns it quickest.

[ad_2]

Accelerating Textual content-to-Video Technology with Calibrated Sparse Consideration

Machine Learning

Dr. Mike

July 22, 2026

Accelerating Textual content-to-Video Technology with Calibrated Sparse Consideration

[ad_1]

Current diffusion fashions allow high-quality video era, however undergo from sluggish runtimes. The massive transformer-based backbones utilized in these fashions are bottlenecked by spatiotemporal consideration. On this paper, we establish {that a} important fraction of token-to-token connections constantly yield negligible scores throughout varied inputs, and their patterns usually repeat throughout queries. Thus, the eye computation in these circumstances could be skipped with little to no impact on the outcome. This statement continues to carry for connections amongst native token blocks. Motivated by this, we introduce CalibAtt, a training-free methodology that accelerates video era through calibrated sparse consideration. CalibAtt performs an offline calibration cross that identifies block-level sparsity and repetition patterns which can be steady throughout inputs, and compiles these patterns into optimized consideration operations for every layer, head, and diffusion timestep. At inference time, we compute the chosen input-dependent connections densely, and skip the unselected ones in a hardware-efficient method. In depth experiments on Wan 2.1 14B, Mochi 1, and few-step distilled fashions at varied resolutions present that CalibAtt achieves as much as 1.58× end-to-end speedup, outperforming present training-free strategies whereas sustaining video era high quality and text-video alignment.

† Tel Aviv College
** Work carried out whereas at Apple

[ad_2]

1...91011...697 Page 10 of 697

<img class="tdb-logo-img" src="https://analyticscampus.com/wp-content/uploads/2025/10/Analytics-Campus-1-300x90.png" alt="Logo" title="" />

<img class="tdb-logo-img" src="https://analyticscampus.com/wp-content/uploads/2025/10/Analytics-Campus-1-300x90.png" alt="Logo" title="" />

<img class="tdb-logo-img" src="https://analyticscampus.com/wp-content/uploads/2025/10/Analytics-Campus-1-300x90.png" alt="Logo" title="" />

A format choice just isn’t a persona take a look at

Consideration and reminiscence

1. You defend sustained focus

2. You want a bodily map of a narrative

3. You care whether or not understanding feels actual

Possession and materials attachment

4. You worth possession, not merely entry

5. You connect reminiscences and identification to things

Why some readers settle for the inconvenience

6. You’ll commerce comfort for a most well-liked expertise

7. You’ve constructed a powerful reader identification

What the choice actually says

The naive baseline this text pushes again on

1. Two regimes for feeding the top-Ok to era

AI suppliers gained’t repair this downside

What’s Gigatoken

The interactive benchmark explorer

Set up

Compatibility mode — drop-in, roughly 200–300× per the creator

Gigatoken API — quickest path, Rust reads the information instantly

Verify your personal tokenizer with out putting in something

Identified points, acknowledged by the creator

Acknowledged AI use

License and packaging

Items of sunfish in every single place

A household meal

Enjoying with their meals

Adopting agentic operations by 4 programs

Agentic 101: Asking for motion

Agentic 201: Guiding outcomes

Agentic 301: Constructing repeatable recipes

Agentic 401: Attaining orchestration

Curriculum method to information agentic adoption

The (unimaginable?) quest for an ethical AI