Research
AutoGen by Microsoft
Microsoft's open-source multi-agent framework, now in maintenance mode
BabyAGI
The 2023 task-loop agent, now a family of open-source experiments
BLOOM
BigScience's 2022 open multilingual model, still downloadable, now a research artefact
ChatDOC
Chat with your documents and click through to the passage behind each answer
Consensus
AI search engine that shows what the research says
Deepnote
Collaborative data notebook with an AI agent and an open file format
Elicit
AI literature search, extraction and systematic reviews with sentence-level citations
Explainpaper
Highlight a confusing line in a paper, get a plain-English fix.
Gemini Notebook (formerly NotebookLM)
Source-grounded research notebooks with audio, video and slide summaries
Genei
UK PDF and web-page summariser, still online but no longer maintained
Guanaco
The 2023 research chatbots that demonstrated QLoRA 4-bit fine-tuning
Ask questions across your files and see the cited passage behind each answer
Hume AI
Voice-AI evaluation and expression measurement
Iris.ai
Enterprise AI harness for R&D documents and research
Kimi 月之暗面
Moonshot AI's assistant and open-weight K3 model with a 1M-token context
Cited AI research, references and writing in one workspace, formerly Afforai
Ad-free Chinese AI search with cited answers and an academic mode
Orca
Microsoft Research's explanation-tuned small models, now a dated research release
Paperguide
AI research workspace for literature search and systematic reviews
Perplexity AI
Answer engine that cites its sources, now with a browser and an agent
Perplexity Pages
Perplexity's cited research pages, retired pending a promised rebuild
Perplexity Pro
Perplexity's $17-a-month tier: model choice, Deep Research and Computer
Perplexity Projects (formerly Spaces)
Perplexity's persistent research workspaces, renamed from Spaces in 2026
Chat with your data, or hand whole analyses to Bloom's agents
Rogo
Agentic AI platform for investment banks, PE and asset managers
Scholarcy
Turns papers into structured flashcards: findings, methods, refs.
Academic search that ranks papers and experts by influence within your field
Semantic Scholar
Free AI-assisted academic search across 200M+ papers, from Ai2
Stanford Alpaca
Stanford's 2023 LLaMA fine-tune that made cheap instruction tuning standard
Stanford's open-source research prototype that drafts cited reports
SWE-agent
Princeton and Stanford's research coding agent, now superseded by mini-swe-agent
Tencent ima (腾讯ima)
Tencent's knowledge-base AI for search, Q&A and writing
Tiangong 天工 (formerly Tiangong AI Search)
Kunlun's former AI search engine, now the 天工 SkyClaw agent workspace
Trinka AI
Grammar and style checking tuned for academic and technical English
WangchanX
VISTEC's open Thai fine-tuning pipeline, models and instruction datasets
WizardLM
The Evol-Instruct model family from Microsoft, inactive since 2024
How many kana there are depends on which list is counted
Japanese is written partly in hiragana and katakana, and a learner's chart shows 46 kana, but the official kan...
EV
19 Oct
A typing-test word is five keystrokes and real English words run longer
Typing tests count a word as any five keystrokes, the space included. Measured over two nineteenth-century nov...
MA
19 Oct
Bolt torque and clamp load across the published nut factors
A tightening torque becomes clamp load through a nut factor, K, that stands in for thread and head friction. O...
KE
19 Oct
Roman numerals on Roman stones fail today's strict standard form
Roman numerals today follow a strict standard form that allows six subtractive pairs, no letter more than thre...
NA
19 Oct
Fine-Particle Air Pollution Linked to Higher Hospital Admission Risk for Up to 21 Days, Study Finds
Researchers matched 20 years of daily admissions in six countries and territories against PM2.5 levels. The ad...
AM
19 Oct
How reading-time estimates turn words and Chinese characters into minutes
A reading-time label divides a text's length by an assumed reading speed, and the speeds in use for English ru...
NA
18 Oct
Neapolitan pizza dough in baker's percentages, from the AVPN and EU rules
Neapolitan pizza dough has two written specifications, the AVPN's and the EU's, and both measure every ingredi...
AM
18 Oct
How the ideal body weight formulas compare from short heights to tall
Four published ideal body weight formulas turn height and sex into one weight, and they were written as drug-d...
AM
18 Oct
CURB-65 death rates checked against the paper that defined the score
CURB-65 grades community-acquired pneumonia in adults with one point each for confusion, raised urea, fast bre...
PR
18 Oct
Most Chinese characters share their four-corner code with other characters
The four-corner code turns the stroke shapes at a Chinese character's corners into digits, a lookup method Wan...
EV
16 Oct
Naegele's rule and the 280-day count give different pregnancy due dates more often than not
A due date is counted from the first day of the last period, either by adding 280 days or by adding nine month...
PR
16 Oct
Riegel's race-time formula comes within about 2% of world records and predicts recreational marathons too fast
Riegel's formula scales a known race time to another distance using an exponent, usually 1.06. Tested against...
NA
15 Oct
Tempo charts disagree on how many beats a minute Andante means
Six published sources give Andante anywhere from 56 to 108 beats a minute, and by our count only 76 and 77 fal...
EV
15 Oct
Glasgow Coma Scale totals hide the eye, verbal and motor scores behind them
A Glasgow Coma Scale total adds three separate observations of the eyes, speech and movement, and the same tot...
PR
15 Oct
Equations estimating a one-repetition maximum match on heavy sets and diverge as reps climb
A one-rep max is the heaviest weight a person can lift once, and published equations estimate it from a lighte...
AM
15 Oct
Dog pregnancy length and why counting from mating widens the due date
A dog's pregnancy averages about 63 days counted from ovulation, but how wide the window is depends on which e...
AM
15 Oct
The usual mid-parental height range missed 12% of Galton's grown children
The usual target height averages the parents' heights, shifts the result 6.5 cm for sex and allows 8.5 cm eith...
PR
15 Oct
Japan's new kanji forms coincide with the mainland's in about a third of pairs
Japan fixed simpler print forms for its everyday kanji in 1949, and the 2010 Jōyō list still prints an old Kan...
EV
15 Oct
Descriptive and algebraic chess notation, and the rules FIDE and PGN files disagree on
Old chess books write moves in descriptive notation, naming each square from the mover's side, while FIDE's La...
NA
15 Oct
An X chromosome comes from fewer ancestors than a family tree allows
A man's X chromosome can come only from his mother's side, while a woman's comes from both parents, so the anc...
NA
15 Oct
Net run rate tables from the ICC add up only when overs are counted in balls
Net run rate is a cricket side's runs per over minus the runs per over scored against it, across every match w...
NA
14 Oct
Humidity adds close to 400 feet to density altitude at Changi
Water vapour is lighter than dry air, so humid air raises density altitude even at a sea-level airport. At Cha...
KE
14 Oct
One study's medium Cohen's d is compatible with effects from below small to large
Cohen's d, Hedges' g and Glass's Δ all divide a difference between two means by a standard deviation, and diff...
NA
14 Oct
Looking up a MAC address by its first three bytes misses the holders of smaller IEEE blocks
A MAC address carries two flag bits and, for a factory address, the organisation that bought its block from th...
PR
14 Oct