Highlight, annotate, and build your AI memory that grows with your learning. Loved by 1M+ users ❤️ We also built YouTube Summary (2M+ users)
San Francisco, CA
- Tweets 6,500
- Following 1,099
- Followers 6,608
- Likes 25,025
Pinned Tweet
Glasp is featured in @Forbes 🙌
With AI-driven features like personalized summaries and AI clones, Glasp is set to transform how we share and build on knowledge. Dive in and see the future of collaborative learning!
8
3
39
9,800
We launched Glasp's Firefox extension on Product Hunt🚀
Glasp is now on @firefox 🦊
Highlight web pages, PDFs, and YouTube transcripts, then summarize them with @claudeai, @ChatGPTapp, or Gemini. Your highlights sync to every browser you sign in to.
We're live on @ProductHunt today. Support us 👇
8
1
15
611
If you highlight in Raindrop.io, those highlights can now live in Glasp too 💧
Connect once and they come across with their notes, colors, and tags. No file export needed.
Only pages you actually highlighted come over, so your plain bookmarks stay where they are.
2
6
327
Here's the tutorial:
glasp.co/posts/how-to-import…
2
94
Glasp retweeted
We just launched Glasp MCP Connector on @ProductHunt 🔌
Connect Glasp to @claudeai and @ChatGPT. Ask "what did I save about deep work?" and your own highlights come back, with sources.
Read-only. Private.
Support us 👇
3
4
9
437
Glasp retweeted
Glasp is finally available on Firefox! 🦊
This release took us much longer than we expected, so we are especially excited to finally share it.
Firefox users can now use @_Glasp to:
• Highlight and annotate web pages and PDFs
• Save and organize ideas in one place
• Add notes and tags to what they read
• Revisit and share what they have learned
• Build a personal AI memory from their highlights
Our goal with Glasp has always been to help people capture valuable ideas while they read, connect them over time, and make their knowledge more useful.
Making Glasp available across more browsers is an important step toward that goal.
A huge thank you to everyone who kept asking us for a Firefox version and patiently waited for the release.
We would love to hear your feedback!
1
3
5
354
Our new research paper is out!
We just released a new research paper: “Language Models Agree With Each Other, Not With Readers”
Many studies have found that language models produce more homogeneous outputs than humans. But the human comparison groups in those studies are usually recruited, instructed, and paid to perform the same task.
We wanted to compare models with people reading naturally, without being told what to find important.
Using Glasp’s public highlighting data, we analyzed:
• 2,523 reader highlight sets
• 120 web documents
• 18 language models
• 11 model vendors
• Models spanning 2024 to 2026
We measured how often two readers or two models selected the same sentences, after controlling for sentence position and length.
The median model pair agreed 2.3 times more than two human readers.
GPT-5.4 and Claude Opus 5, despite coming from rival labs, agreed 5.1 times more than two readers.
We also tested four models released after the original analysis was completed:
• GPT-5.5
• GPT-5.6 Luna
• GPT-5.6 Sol
• GPT-5.6 Terra
This gave us an out-of-sample test of the paper’s main finding.
None of the four models agreed with readers significantly more than human readers agreed with each other. None surpassed the best model in the original panel or approached the estimated crowd-consensus ceiling.
But their agreement with the original panel’s frontier models was extremely high.
Across 12 comparisons, the new models reached a median agreement of +0.224 with the panel’s frontier incumbents. That is higher than the +0.203 agreement between GPT-5.4 and Claude Opus 5.
The newer models did not move clearly beyond the human agreement level.
They did, however, continue to converge strongly with other frontier models.
Another result surprised us.
Newer models are becoming more aligned with human readers, but they are becoming even more aligned with one another. From 2024 to 2026, agreement with readers increased 2.9 times, while agreement between models increased 3.2 times.
Models are becoming more human-like and more alike at the same time.
We also tested whether models simply preferred a different style of sentence. After controlling for sentence length and position, none of the surface features we measured reliably separated model-selected sentences from reader-selected sentences.
Models and readers choose sentences that look similar, but they choose different ones.
This matters for search, summarization, recommendations, research, AI evaluation, and any system that uses multiple models as if they represent independent perspectives.
Using several frontier models may create the appearance of diversity without providing genuinely different views of what matters.
This work was conducted with @KeiWatanabe17 at @_Glasp
2
1
3
305
Glasp retweeted
We just released a new research paper: “Language Models Agree With Each Other, Not With Readers”
Many studies have found that language models produce more homogeneous outputs than humans. But the human comparison groups in those studies are usually recruited, instructed, and paid to perform the same task.
We wanted to compare models with people reading naturally, without being told what to find important.
Using Glasp’s public highlighting data, we analyzed:
• 2,523 reader highlight sets
• 120 web documents
• 18 language models
• 11 model vendors
• Models spanning 2024 to 2026
We measured how often two readers or two models selected the same sentences, after controlling for sentence position and length.
The median model pair agreed 2.3 times more than two human readers.
GPT-5.4 and Claude Opus 5, despite coming from rival labs, agreed 5.1 times more than two readers.
We also tested four models released after the original analysis was completed:
• GPT-5.5
• GPT-5.6 Luna
• GPT-5.6 Sol
• GPT-5.6 Terra
This gave us an out-of-sample test of the paper’s main finding.
None of the four models agreed with readers significantly more than human readers agreed with each other. None surpassed the best model in the original panel or approached the estimated crowd-consensus ceiling.
But their agreement with the original panel’s frontier models was extremely high.
Across 12 comparisons, the new models reached a median agreement of +0.224 with the panel’s frontier incumbents. That is higher than the +0.203 agreement between GPT-5.4 and Claude Opus 5.
The newer models did not move clearly beyond the human agreement level.
They did, however, continue to converge strongly with other frontier models.
Another result surprised us.
Newer models are becoming more aligned with human readers, but they are becoming even more aligned with one another. From 2024 to 2026, agreement with readers increased 2.9 times, while agreement between models increased 3.2 times.
Models are becoming more human-like and more alike at the same time.
We also tested whether models simply preferred a different style of sentence. After controlling for sentence length and position, none of the surface features we measured reliably separated model-selected sentences from reader-selected sentences.
Models and readers choose sentences that look similar, but they choose different ones.
This matters for search, summarization, recommendations, research, AI evaluation, and any system that uses multiple models as if they represent independent perspectives.
Using several frontier models may create the appearance of diversity without providing genuinely different views of what matters.
This work was conducted with @KeiWatanabe17 at @_Glasp
2
5
10
933
Honored to be featured and to share some of what we’ve learned building Glasp over the past 5 years. Thank you for the great article!
【記事更新】凄まじい行動量でアメリカで事業を立ち上げ、世界350万ユーザーに到達した「Glasp」さんに、持続可能な形でグロースするコツなど取材しました✍️
●1. 最初のユーザーは創業者が集める
最初の100人は友人や元同僚にDMなど、1,000人まではLinkedInやSlackのコミュニティで会話しながら。初期は創業者が営業のように動くことが最も効率的だった。
●2. 800回のオンボーディングで改善
ユーザーと通話をしながら画面共有をして使ってもらうと「躓きポイント」が可視化された。価値の高い発見は「ユーザーが言ったこと」よりも「ユーザーが見せてくれたもの」の中にあった。
●3. ユーザーの言葉から価値を見つけた
「Glaspをどんなサービスだと理解しましたか？」と聞くとそこに一番のバリューが現れていた。これを何度も聞いていくと「Glaspとは何か？誰のためのものか？」がだんだんわかってきた。
●4. 複利が効くSEOと口コミに集中した
獲得コストをほぼゼロに抑えて長期で効果が積み上がるチャネルを選択した。数百本の記事が、数年後には毎月数万人の新規ユーザーを生む集客資産になった。大規模な資金調達なしでも数百万ユーザーに。
●5. 不完全でもトレンドに早く乗った
拡張機能の「YouTube Summary with ChatGPT」をリリース。14カ月で100万インストールを突破。画面にGlaspのロゴを入れることでSNSにスクショが投稿されるたびに認知が広がる仕組みを作れた
●6. パワーユーザーをβテストに巻き込む
熱心なユーザーを新機能のテストに招待するとバグ発見だけでなくユーザーとの関係性が強くなりコミュニティが強化された。
●7. 訪問数ではなく継続利用を指標にした
重要指標は「初回から1ヶ月以内に2回目のハイライトをした割合」と「月1回以上ハイライトした人（MAU）」。月4〜5回使うユーザーは「半年後の継続率」が約2倍も高かったため指標に組み込んだ。
●8. 共有動機を持ったユーザーを狙う
当初はPM向けと考えていたが、編集者と情報を共有する「ライター」のほうが相性がよいとわかった。口コミで広げたいなら共有する動機を持っているユーザーに使ってもらうのが大事だった。ぴったりな層は隣接領域にいることがある。
●9. ChatGPTからの1日の流入が37倍に
AIからのログを分析し情報量を増やし、タイトルを質問形式にした。AIがアクセスしていた404を「未充足の需要」と捉えてページを作成。ChatGPT経由での1日あたりのセッションが37倍に成長した。
●10. シリコンバレーから発信した
新しい技術やサービスに世界中の注目が集まる場所で活動することで情報が世界へ広がりやすくなる。
【👇詳細記事】
約800回のオンボーディングで連続改善。低コスト×複利効果で持続的にグロース。世界350万人がつかう「Glasp」が実践した10のプロダクトとマーケの施策。
markelabo.com/n/n7acf7a1c926…
1
4
436
Glasp retweeted
Feeling grateful to be featured by such a well-respected media outlet in Japan. Thank you for taking the time to dive so deeply into our story and learnings over the past 5 years. We’ll keep building toward our mission and vision. 🙌
【記事更新】凄まじい行動量でアメリカで事業を立ち上げ、世界350万ユーザーに到達した「Glasp」さんに、持続可能な形でグロースするコツなど取材しました✍️
●1. 最初のユーザーは創業者が集める
最初の100人は友人や元同僚にDMなど、1,000人まではLinkedInやSlackのコミュニティで会話しながら。初期は創業者が営業のように動くことが最も効率的だった。
●2. 800回のオンボーディングで改善
ユーザーと通話をしながら画面共有をして使ってもらうと「躓きポイント」が可視化された。価値の高い発見は「ユーザーが言ったこと」よりも「ユーザーが見せてくれたもの」の中にあった。
●3. ユーザーの言葉から価値を見つけた
「Glaspをどんなサービスだと理解しましたか？」と聞くとそこに一番のバリューが現れていた。これを何度も聞いていくと「Glaspとは何か？誰のためのものか？」がだんだんわかってきた。
●4. 複利が効くSEOと口コミに集中した
獲得コストをほぼゼロに抑えて長期で効果が積み上がるチャネルを選択した。数百本の記事が、数年後には毎月数万人の新規ユーザーを生む集客資産になった。大規模な資金調達なしでも数百万ユーザーに。
●5. 不完全でもトレンドに早く乗った
拡張機能の「YouTube Summary with ChatGPT」をリリース。14カ月で100万インストールを突破。画面にGlaspのロゴを入れることでSNSにスクショが投稿されるたびに認知が広がる仕組みを作れた
●6. パワーユーザーをβテストに巻き込む
熱心なユーザーを新機能のテストに招待するとバグ発見だけでなくユーザーとの関係性が強くなりコミュニティが強化された。
●7. 訪問数ではなく継続利用を指標にした
重要指標は「初回から1ヶ月以内に2回目のハイライトをした割合」と「月1回以上ハイライトした人（MAU）」。月4〜5回使うユーザーは「半年後の継続率」が約2倍も高かったため指標に組み込んだ。
●8. 共有動機を持ったユーザーを狙う
当初はPM向けと考えていたが、編集者と情報を共有する「ライター」のほうが相性がよいとわかった。口コミで広げたいなら共有する動機を持っているユーザーに使ってもらうのが大事だった。ぴったりな層は隣接領域にいることがある。
●9. ChatGPTからの1日の流入が37倍に
AIからのログを分析し情報量を増やし、タイトルを質問形式にした。AIがアクセスしていた404を「未充足の需要」と捉えてページを作成。ChatGPT経由での1日あたりのセッションが37倍に成長した。
●10. シリコンバレーから発信した
新しい技術やサービスに世界中の注目が集まる場所で活動することで情報が世界へ広がりやすくなる。
【👇詳細記事】
約800回のオンボーディングで連続改善。低コスト×複利効果で持続的にグロース。世界350万人がつかう「Glasp」が実践した10のプロダクトとマーケの施策。
markelabo.com/n/n7acf7a1c926…
1
1
4
507
Thank you for featuring Glasp!
You’re wasting Chrome’s real power.
Most people browse like it’s 2010.
Meanwhile, others automate half their work.
These 10 tools change everything.
1/ Glasp
Highlight anything on the web and save it straight to your personal knowledge base.
Articles, PDFs, insights - all in one place.
2/ Merlin
An AI assistant on every website.
No more switching tabs to ChatGPT. Just click and ask.
3/ Tango
Click through any process once…
It instantly turns it into a step-by-step guide with screenshots.
4/ Scribe
Records your screen actions and converts them into a clean how-to document.
Zero manual writing.
5/ Bardeen
Automates repetitive browser tasks.
Scraping, form filling, moving data - no code needed.
6/ Eightify
Turns long YouTube videos into key takeaways in seconds.
Perfect for learning fast.
7/ Tactiq
Live transcribes your Google Meet and Zoom calls.
Automatically saves notes so you never miss anything.
8/ GoFullPage
Takes full-page screenshots of any website in one click.
No stitching. No hassle.
9/ Wordtune
Highlight any sentence and rewrite it instantly.
Cleaner, sharper, more professional.
10/ Magical
Creates text shortcuts that auto-fill anywhere online.
Save hours on repetitive typing.
Save this before everyone else finds out.
4
343
📊 Quick poll: Desktop or mobile?
We're curious: where do you use Glasp's YouTube Summary the most? Desktop, mobile, or both? It takes one tap to answer, and your vote directly shapes what we improve next.
100%
Desktop (extension)
0%
Mobile app
0%
Both
0%
There's a mobile app??
3 votes • Final results
2
4
250
Glasp retweeted
For a few years now we have been mapping where individuality lives in reading behavior. The question we had never asked is the one that turned out to matter most: how long does it last?
Quick background. On a social highlighter like @_Glasp, the strongest personal signal is not where you highlight inside a document. The crowd dominates that. It is which documents you choose to read at all. A reader's own history identifies their documents far better than a matched stranger's. But every result we had was a snapshot. So we asked: is that signature a trait, or a state?
The distinction is not academic. Production recommenders bet on state. They weight recency, chase fresh data, and forget the old. If reading taste is a rolling state, that is the right call. If it is a trait, you are quietly discarding a durable asset.
So we froze each reader's first six months as a profile and watched it age against their later choices, out past two years. It barely decays. At a 6 to 12 month gap, retention is essentially 1.00. We found no statistically detectable decline through two years.
Three things still surprise me:
1. A profile built from a reader's earliest documents, around 20 months old on average, still ranks their next reads about 3x better than any simple non-personal baseline.
2. Popularity ties with random at the individual level (0.229 vs 0.227). Because one person's reading lives in the long tail, "what is popular" tells you almost nothing about what they read next.
3. It is not just "they keep visiting the same sites." Roughly 90% of the advantage survives even after we remove every source the profile had ever seen.
The takeaway for anyone building personalization: aggressive forgetting is a mistake on this kind of signal. Recency weighting earns its keep on short-term intent, not by replacing the durable base underneath. Old preference data does not lose its identity value.
Reading identity, it turns out, is closer to who you are than to what you did last week.
Paper (with @KeiWatanabe17)
2
4
7
476
Glasp retweeted
Every recommendation system has a cold-start problem. Ours is unusually concrete.
On @_Glasp, the signal we care about most is crowd salience: which passages many readers independently choose to highlight. It is powerful, but it only exists once a document has been read. A new article starts at zero.
Our new paper asks whether we can close that gap and predict a document's crowd highlights from its text alone, before any marks accumulate.
The result is really a study in calibration. A trained logistic ranker beats the strong lead (position) baseline by a small but robust +0.044 average precision, pre-registered and bootstrapped. Off-the-shelf LLMs and generic extractive summarizers do not beat lead at all.
For anyone building reading or knowledge tools: position is a surprisingly hard baseline to beat, and the real modeling gains live in the long tail, not the front page. Useful to know before you over-engineer.
With my co-first author @KeiWatanabe17.
#MachineLearning #InformationRetrieval #BuildingInPublic
1
2
7
335
Glasp retweeted
"Popular highlights" treat the crowd as one consensus. It isn't.
Paper III of our research series at @_Glasp is now on arXiv: "Factions Within, Uncertain Across: Within-Document Reader Sub-Groups in Social Highlighting" (co-authored with @KeiWatanabe17).
When many readers highlight the same article, platforms average them into a single ranking of what matters. We asked whether that average hides structure. It does.
Within a document, readers form strong sub-groups. Pairs agree with each other far beyond what shared salience, mark density, and sentence popularity predict (z = +6.3, significant in 88% of documents). Shared section engagement explains only about 40% of the excess. The majority is finer: readers marking the same sentences within a region, more than chance allows.
The contrast with our earlier papers makes this striking. The individual within-document signal is a whisper (a person's own history beats a stranger's by just +0.017 average precision). The group signal, in the exact same place, is loud. A quiet individual inside a strongly factional crowd.
The honest part: whether the same readers keep grouping together across documents is unresolved. The test is underpowered at current co-readership density, estimates are positive but never significant, and we say so plainly rather than overclaiming in either direction.
Product takeaway: documents with existing readers could show multiple "ways people marked this" instead of one map. Portable reader segments remain an open question.
2
2
6
458
Glasp retweeted
We just published our second research paper, built on Glasp data.
When we started @_Glasp, the dream was simple: if we understand what each person highlights, we can personalize their entire reading experience.
This paper is us testing that dream rigorously. The answer surprised us.
Your reading history does say a lot about WHICH articles are yours. With a clean, leakage-free test, we could identify a person's documents among their co-readers' choices, even when the topics matched.
But WITHIN a document? Personalization stopped working. A model that knew your entire highlight history could not beat the shared, impersonal sense of what matters. Even frontier LLMs lost to a simple lead baseline at predicting highlights.
Our conclusion: personalization lives at the selection layer, not the salience layer. People differ in what they read. What stands out is mostly shared.
And honestly, my favorite part: we found a bias in our own evaluation that inflated our first result, audited it, and published the corrected number instead. That is the kind of research we want to do.
So maybe the future is not personalizing each reader harder. It is aggregating readers, turning shared salience into collective intelligence.
Co-authored with my co-founder @KeiWatanabe17.
#Glasp #Research #Personalization #ReadingTech
2
2
8
1,445
Glasp retweeted
What does a highlight say about you?
On @_Glasp, people mark the sentences that matter to them as they read. Over time those marks pile up into something quiet and personal: a record of where your attention landed. We wanted to know how much of "you" is actually in there.
So we asked a question that sounds simple and turns out not to be. If I show you and a stranger the same article, does your own reading history predict what you will highlight better than the stranger's does?
Because many people on Glasp highlight the same pages, we could hold the text fixed and measure this cleanly, separating what is generally salient (anyone notices it), what the crowd marks, and what is left over as personal.
The first answer was humbling. Highlighting is social. What you mark in a given document is predicted far better by the crowd than by any model of you. We even built an LLM "twin" from a person's own highlighting history and put it up against a simple, well estimated crowd. The crowd won. Inside a single document, the part that is uniquely you is, at most, a whisper.
That could read as a deflating result. It is the opposite. Individuality does not vanish. It moves.
When we changed the question from "what is important in this text" to "which of the already important passages are yours to mark," a person's history suddenly became a strong, clean predictor. You are not very legible in what you find salient. You are very legible in what you select. And that selection is mostly stable, thematic taste: the kinds of things you keep returning to across everything you read.
I find that quietly beautiful. It suggests that the signature of a reader is less about seeing different things on the same page, and more about which pages, and which threads, a person chooses to lean toward. The self shows up in attention and curiosity, not in disagreement over what a paragraph means.
One methodological note we care about. A lot of "personalization" results leak: the thing you are trying to predict slips into the profile you predict it from. When we closed those leaks, used a dense crowd, and made the comparison fair and model matched, the honest picture above is what remained.
Deep thanks to @KeiWatanabe17, and to everyone whose highlights make Glasp a place where a question like this can even be asked.
2
3
6
755
Leverage your Kindle highlights with our API :)
You can now export your Kindle highlights via the Glasp API 📚
Your highlights, yours to take anywhere.
1. Import your Kindle highlights with our Chrome Extension (manual or auto-sync)
2. Export them via the Glasp API
And as you import, Glasp emails you a daily highlight review so you can resurface what you've read :)
3
193
To download and import Kindle highlights to Glasp, please check this tutorial.
How to Download Highlights and Notes from Kindle
glasp.co/posts/how-to-downlo…
3
181