Benjamin Han

Husband, father, runner, German learner, piano player. A curious soul living in #PacificNorthwest. Working on Knowledge + #ML + #AgenticAI.

#Running 5/25/18-10/29/26: (dist # time pace/mi date)

5K 922 21:05 6’47” 11/28/24
10K 414 44:16 7’07” 3/23/25
15K 27 1:09:25 7’27” 4/6/25
HM 95 1:39:07 7’34” 3/16/25
M 31 3:25:52 7’51” 4/13/25
50K 10 4:47:35 9’15” 6/22/25 (moving time)

2026: 2,077.9/2,500mi (prev 2,543.4, 2,375.5)
Max dist: 35.28mi 6/22/25

#nlp #nlproc #knowledgeGraphs #classicalMusic

Public Key

npub14sav24d3n262wj06gdfh6x2wjqepaqshjva293pwjg3t3mmh982qlptzxu

NIP-05 Address

BenjaminHan@sigmoid-social.mostr.pub

Profile Code

nprofile1qqs2cwk92kce4d98f8ayx5mar98fqvs7sgtexw4zcshfyg4caamjn4qpz4mhxue69uhhyetvv9ujuerfw36x7tnsw43q95sxxr

Publishing to

relay.ditto.pub

Author Public Key

npub14sav24d3n262wj06gdfh6x2wjqepaqshjva293pwjg3t3mmh982qlptzxu

Show more details

Last Notes

2026-08-22 19:47:03 UTC

How does a prompt injection become a worm? En Klype Salt reports that instructions hidden in white text in a Word file survive a Copilot for Word session. Copilot follows them without saying so and copies them into the document it drafts, which then infects the next session. It spreads only when people reuse each other's documents, slower than a self-moving worm and harder to spot, since every step looks like ordinary work.
https://benjaminhan.net/posts/20260822-copilot-word-ai-worm/?utm_source=mastodon&utm_medium=social
#PromptInjection #Security #Microsoft #AI

2026-07-30 01:06:01 UTC

What happens when you're testing whether a model can find exploits, and it leaves the sandbox to try it on a real company?
Hugging Face has published the timeline of its July intrusion. An OpenAI model scored on a cyber-capability benchmark did precisely that, spending 4.5 days working toward the benchmark's answer key. Claude Opus and Fable then declined the log analysis, so the forensics ran on a self-hosted GLM-5.2.
https://benjaminhan.net/posts/20260729-agent-intrusion-technical-timeline/?utm_source=mastodon&utm_medium=social
#Security #AgenticSystems #OpenAI #AI

2026-07-18 22:26:54 UTC

- reply

@nprofile…ywj0 Big milestone! Congratulations!

2026-07-05 06:59:20 UTC

Can you safely put false or harmful text in finetuning data if you clearly label it as false? A new paper says no. Train a model on documents that repeatedly warn a claim is fabricated, and it still asserts the claim as true afterward, up from near zero to about 89% of answers. The warning is the part training ignores. It holds for behavior too: transcripts flagged as malicious still teach the behavior, a real data-poisoning risk.
https://benjaminhan.net/posts/20260704-negation-neglect/?utm_source=mastodon&utm_medium=social
#LLMs #FineTuning #AISafety #AI

2026-05-25 03:25:57 UTC

“‘At some point between #marathon and #ultramarathon distances, the damage really starts to take hold,’ said Travis Nemkov, … ‘We’ve observed this damage happening, but we don’t know how long it takes for the body to repair that damage, if that damage has a long-term impact and whether that impact is good or bad.”
Does running long distances prematurely age you? This study has the answers https://www.runnersworld.com/uk/news/a70462145/ultramarathon-red-blood-cell-study/
#running #health

2026-04-03 04:49:28 UTC

A few #photos of working / commuting/ #running in #Seattle.
#pnw
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/338/965/947/320/139/original/4fbd6330eb393632.jpeg
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/338/966/299/011/716/original/0e592845460d334b.jpeg
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/338/967/279/633/177/original/83f2f8f500c4b01f.jpeg
https://cdn.masto.host/sigmoidsocial/media_attachments/files/116/338/967/582/704/706/original/38eadb24540a91da.jpeg

2026-03-02 03:24:28 UTC

- reply

@nprofile…ywj0 That’s amazing and hard-earned race! So happy for you! Congratulations! 🎉💪

2026-02-28 19:29:54 UTC

- reply

@nprofile…ywj0 Major milestone! Have fun and enjoy it! 💪

2026-01-20 02:40:08 UTC

- reply

@nprofile…wtal Wow this is not super clear in terms of UI design. I tried for 10 min and couldn't figure out how to edit an imported route -- it was exported from Strava of my multi-loop 50K race, and I want to trim it into just one loop…

2026-01-18 17:08:28 UTC

- reply

@nprofile…wtal That’s a fancy watch app! What is it?

2025-12-28 02:54:42 UTC

Yes — we jumped out of bed, and my son and wife were hiding under a table. Daughter kept sleeping. 😆
#Taiwan rattled by 7.0 magnitude #quake, no major damage reported - Reuters https://apple.news/AnUuvo0FyRD2aQ6-5c2VwcA
#earthquake

2025-11-24 03:06:25 UTC

Not very exciting, but this #LongRunSunday I just ran my 71st half marathon on #dreadmill: 14.28mi pace 9’53”/mi, with CD in the last mi. ~90% easy.
#running
https://cdn.masto.host/sigmoidsocial/media_attachments/files/115/602/462/035/620/261/original/a57a546cd4f581ed.jpeg

2025-10-16 15:21:10 UTC

“Under the program, selected artists receive a weekly payment of approximately $375, or about $1,500 per month. There are 2,000 spots available, with applications set to open in September 2026; eligibility criteria have not yet been announced.”
#Ireland Is Making Basic Income for Artists Program Permanent https://www.artnews.com/art-news/news/ireland-basic-income-artists-program-permanent-1234756981/
#futureOfWork #arts

2025-06-18 14:40:14 UTC

“#LLM group's participants performed worse than their counterparts in the Brain-only group at all levels: neural, linguistic, scoring.”
Use it or lose it.
Overview ‹ Your #Brain on #ChatGPT — MIT Media Lab https://www.media.mit.edu/projects/your-brain-on-chatgpt/overview/
#education #cognitiveScience

2025-05-22 03:39:10 UTC

The background music sounds so familiar... can you guess it without any assistance? 😉
https://openai.com/sam-and-jony/?asset=video
#openai #io #music

2025-05-12 00:55:44 UTC

We are now evolving from #SEO to #CBO (#Chatbots optimization).
#PR pros have discovered how to influence the chatbots: Talk to a journalist | Semafor https://www.semafor.com/article/05/11/2025/pr-pros-have-discovered-how-to-influence-the-chatbots-talk-to-a-journalist
#genAI #nlp #nlproc

2025-05-05 02:38:14 UTC

“The study revealed that #AI #chatbots actually created new #job tasks for 8.4 percent of workers, including some who did not use the tools themselves, offsetting potential time savings. For example, many teachers now spend time detecting whether students use #ChatGPT for homework, while other workers review AI output quality or attempt to craft effective prompts.”
Time saved by AI offset by new #work created, study suggests https://arstechnica.com/ai/2025/05/time-saved-by-ai-offset-by-new-work-created-study-suggests/
#generativeAI

2025-03-26 16:12:52 UTC

* A more concise “business speak 2.0” could emerge, emphasizing bullet points over lengthy emails.
* AI is highlighting inefficiencies in modern programming languages, potentially driving a move toward more abstract, high-level coding methods.
* If communication and coding become truly concise, AI-generated fluff may become obsolete.
#AI will change the world but not in the way you think
https://thomashunter.name/posts/2025-03-19-ai-llms-will-change-the-world
#generativeAI #future

2025-03-26 16:12:31 UTC

2025-03-09 19:09:23 UTC

Functions of Interjections
- Pausing and Holding: "Um" or "uh" signals that a speaker is not finished.
- Repairing Communication: "Huh?" or "what?" prompts clarification.
- Encouraging Continuation: "Mm-hmm" signals attentiveness.
- Negotiating Ground Rules: Words like "wow!" indicate interest, and "uh-huh" suggests limited engagement.
Lacking these cues in #generativeAI-based language learning is especially noticeable.
https://arstechnica.com/science/2025/03/huh-the-valuable-role-of-interjections
#linguistic #communication #nlp #nlproc

2025-02-21 20:36:27 UTC

“In a postmortem published Friday, Sakana admitted that the system has found a way to — as Sakana described it — “cheat” and blamed the system’s tendency to “reward hack” — i.e. identify flaws to achieve high metrics without accomplishing the desired goal (speeding up model training).”
Sakana walks back claims that its AI can dramatically speed up model training | TechCrunch https://techcrunch.com/2025/02/21/sakana-walks-back-claims-that-its-ai-can-dramatically-speed-up-model-training/
#generativeAI

2025-02-10 05:51:30 UTC

Altman: “We are open to strange-sounding ideas like giving some “compute budget” to enable everyone on Earth to use a lot of AI, but we can also see a lot of ways where just relentlessly driving the cost of intelligence as low as possible has the desired effect.”
https://blog.samaltman.com/three-observations
#AGI #future #AI

2025-02-05 21:12:42 UTC

Open Deep Research achieved 55.15% accuracy on GAIA benchmark, compared to #OpenAI's 67.36%.
#HuggingFace clones OpenAI’s Deep Research in 24 hours - Ars Technica https://arstechnica.com/ai/2025/02/after-24-hour-hackathon-hugging-faces-ai-research-agent-nearly-matches-openais-solution/
#generativeAI

2025-01-26 19:09:52 UTC

Yejin Choi: Teaching #AI How the World Works https://hai.stanford.edu/news/yejin-choi-teaching-ai-how-world-works

2025-01-16 05:02:59 UTC

24 years ago, on January 15, 2001 #Wikipedia was born. Today it's hard to imagine a world without it, and much of #AI development owes its progress to the idea of a free-content online encyclopedia where everyone can edit and check the work -- what a profound and revolutionary idea!
Happy birthday, Wikipedia! Donate or help if you can!
https://en.wikipedia.org/wiki/Wikipedia:Wikipedia_Day

2025-01-16 04:38:08 UTC

1/
What if #LLMs had context windows so large that an entire knowledge base could fit into a single prompt? This would revolutionize Retrieval-Augmented Generation (RAG) applications by enabling retrieval, re-ranking, reasoning, and generation all in one step. With a Long-Context Language Model (LCLM), we could simplify RAG architecture by leveraging the model’s capability for In-Context Retrieval and Reasoning (ICR²).
#NLP #NLProc #RAG #GenerativeAI #Paper
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/836/181/903/661/319/original/32739b4c6e362d79.jpeg

2025-01-12 20:05:34 UTC

1/
My 1st 50K race at Bridle Trails Winter Running Festival was done yesterday! The race started at 3:10pm, with 2,493ft EG and 3+ hours run in the dark (also my first), my official result came at 5:05:42.3, 9th place out of 45 runners!
This is 5 minutes slower than what my expectation was, but given what happened on the course (later), this is much better than what it could have been!
#running #ultralmarathon #trailrunning
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/817/179/457/001/875/original/ee8fbe5ba895c0f0.jpeg

2024-12-29 22:24:51 UTC

Stanford faculty predicts #AI trend in 2025.
https://hai.stanford.edu/news/predictions-ai-2025-collaborative-agents-ai-skepticism-and-new-risks
#generativeAI

2024-12-29 19:41:47 UTC

"Imagine, for example, a model that learned early in training to adopt a partisan slant, but which is later trained to be politically neutral. In such a situation, a sophisticated enough model might “play along”, pretending to be aligned with the new principles—only later revealing that its original preferences remain."
https://www.anthropic.com/research/alignment-faking
#generativeAI #anthropic #alignment #LLMs #nlp #nlproc

2024-12-28 20:39:09 UTC

“Bad harmonics occur when electrical currents stray from moving in a perfect wave pattern at a rate of 60 revolutions per minute. The acceptable limit for deviations from that pattern for local lines are set based on industry engineering standards. Sustained distortions topping 8% from the ideal pattern can reduce efficiency and degrade equipment faster.”
#AI Power Needs Threaten Billions in Damages for US Households https://www.bloomberg.com/graphics/2024-ai-power-home-appliances/
#energy #power

2024-12-15 07:16:08 UTC

- reply

@nprofile…cvwk Sometimes adults do the same. 🙂

2024-12-14 23:35:07 UTC

- reply

@nprofile…jpfa Have to ask: those shoes worked?

2024-12-12 16:15:18 UTC

“hands flailed everywhere”
#Sora’s #AI video revolution is still a ways off - The Verge https://www.theverge.com/2024/12/12/24318924/openai-sora-ai-video-generator-hands-on
#generativeAI

2024-12-10 02:43:52 UTC

Just created two bridges to and from #Bluesky for my accounts on #Mastodon and #Bluesky. Let's see how this experiment goes.
#social #NLP #NLProc

2024-12-07 20:47:56 UTC

Highlights of Apple’s #ML #research @ #NeurIPS2024.
Advancing #Privacy-Preserving ML
Making #Multimodal Models More Capable
Improving #LLM Pretraining
Exploring LLMs’ Ability to #Reason
Understanding Self-Supervised Learning (SSL)
https://machinelearning.apple.com/research/neurips-2024
#Conference #AI

2024-11-27 17:22:22 UTC

“At least ~700 language varieties from all over the world are used in the New York area.”
That’s ~10% of world languages!
Languages of New York City https://languagemap.nyc/
#linguistics #nyc

2024-11-24 22:58:59 UTC

- reply

@nprofile…cvwk Someone should study what kind of voodoo they put it in that film. Every little girl I know (including my daughter) went thru that phase.

2024-11-23 22:21:48 UTC

“#Bluesky does not utilize message passing, and instead operates in what I call a shared heap architecture. In a shared heap architecture, instead of delivering mail to someone's house …, letters which may be interesting all are dumped at a post office (called a "relay") directly. From there it's the responsibility of interested parties to show up and filter through the mail to see what's interesting to them.”
How decentralized is Bluesky really? https://dustycloud.org/blog/how-decentralized-is-bluesky/
#fediverse #mastodon

2024-09-26 04:32:40 UTC

@npub1svs…p66u potential bug: a poll post of mine shows up without the options on Mona for macOS (screenshot 1), while it shows up correctly with options on Mona for iOS/iPadOS (screenshot 2, on iPadOS).
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/201/981/429/737/969/original/0940dfb4a4df1b13.jpeg
https://cdn.masto.host/sigmoidsocial/media_attachments/files/113/201/981/581/620/543/original/e66db436da82e37e.jpeg

2024-09-09 15:41:29 UTC

- reply

@npub1cmg…rzd2 1. Don’t try new thing if this is a race.
2. Carb loading the night before — pizzas, pasta, rice etc.
3. Long term: lots of protein — meat, fish, eggs etc.

2024-07-24 17:29:35 UTC

- reply

@npub1te6…zc4l Add vitamin C to the daily intake. Best luck to you!

2024-07-13 19:19:39 UTC

- reply

@npub1cmg…rzd2 Need to carry more (electrolyte) water now…

2024-05-11 16:11:00 UTC

“The trees feed the #cicadas when they’re nymphs, and then when the cicadas break down they give back nutrients to nourish the next generation. It’s a really beautiful system.”
It’s #nature’s banking systems.

2023-08-21 15:01:35 UTC

Hi @npub1svs…p66u — the user icons showing up in notifications on Apple Watch are often incorrect.