iiinova — where your digital self keeps evolving

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 01/10/2026

September #running recap: fewer but faster miles. 251.7 → 217.0mi, pace 9'44" → 9'32"/mi. I'm at 1,882 of 2,500mi for the year, now 11.7mi *ahead* of plan (yay)! The month opened w/ the Redmond Harvest Half, then a planned recovery week. VO₂max hit an all-time high of 55.47, and the longest run […]

sigmoid.social

Original post on sigmoid.social

022

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 28/09/2026

#Seattle claims the No. 4 city in the world with an overall score of 92.8, according to Oxford Economics’ 2026 Global Cities Index. The top 5: New York City, London, Paris, Seattle, and San Francisco. www.oxfordeconomics.com/global-citi… #city

oxfordeconomics.com

044

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 28/09/2026

Last 20-miler before Snohomish, and my HM #94: 14mi capped at HR 142, then 6 capped at 149. 8'32"/mi, HR 144/159. My phone unlocked itself in my vest pocket, chopped my laps into few-second splits and killed my 180bpm playlist at ~14mi (sorry for the F-bomb […] [Original post on sigmoid.social]

Deep blue sky over the Marymoor R/C Club's Edward B. Sand Memorial Airfield in Redmond, a small model plane flying high in the middle of the frame, picnic tables and a white clubhouse along the fence, and trees starting to turn at the edge of the field.

Route map of a 20-mile run from Marymoor Park, south along Lake Sammamish and north up the Sammamish River Trail, the track colored by heart rate; below it, heart rate and pace against distance, flat near 141 bpm for 14 miles, then stepping up to about 150 for the last 6.

Per-mile splits table for the 20 miles, pace 8:29 to 8:46 for miles 1 to 14 and 8:12 to 8:31 for miles 15 to 20, above the coach's notes on the steady first 14 miles, the quiet calf and the new fuel pouch.

021

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 27/09/2026

Can we read what #LLMs are thinking but not saying? A paper from #Anthropic finds a small set of word-linked patterns inside #Claude that it can report, control on request, and use to reason silently, like the global workspace theory in #neuroscience. How do these findings connect to […]

sigmoid.social

Original post on sigmoid.social

003

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 26/09/2026

#CoachMiles prescribed an easy #parkrun today (a 20-miler is waiting for me tomorrow), and it turned out to be my 900th 5K run! I behaved for the first 2.4mi (9'14"/mi), but couldn't help but splurge on the last 0.7mi at 7'11"/mi, finishing at 6'37"/mi! #running #5K #pnw #photo

Start of the Redmond Central Connector Trail parkrun: a smiling volunteer in a blue parkrun vest stands by a yellow START sign under sunlit trees while runners gather on the path

The giant mosaic REDMOND letters on a plaza in downtown Redmond, backlit by a bright morning sun in a clear blue sky, their shadows stretched across the pavement

Route map of the 3.15-mile parkrun on the Redmond Central Connector Trail, the track colored by heart rate from green to red; below it, heart rate and pace against distance, both flat through 2.4 miles and then climbing to HR 165 and 6:38/mi pace at the finish

Per-mile splits table: 9:13, 9:21, 8:06, then 6:37 for the last 0.14 mile, with HR 133, 136, 147, 164; below it, CoachMiles' notes on the capped first 2.4 miles, the quick finish, and a third quiet calf run in a row

0101

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 21/09/2026

Does routing to whichever model says it can answer lead to more success? 20 frontier models were asked on six benchmarks whether they could answer w/o help. They mostly agree about which questions are hard, but models saying yes don't necessarily answer more correctly. Counting hedge words in a […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 21/09/2026

This #LongrunSunday I ran #HM #93 as a rehearsal of the coming Snohomish River Run #Marathon, HR capped at 142. I got 13.31mi at 8'39"/mi, HR 140/149. Fastest mile was the last one, 8'20". Decoupling was 3.0% and efficiency factor was 1.320 m/heartbeat, same […] [Original post on sigmoid.social]

Selfie in a sleeveless running top, orange cap and sunglasses, hydration vest on, standing in front of a large painted Earl Winehart mural in Snohomish.

A timber archway over a paved path reading City of Snohomish Centennial Trail, with trees and blue sky behind it.

Route map with HR-colored track along the Centennial Trail; HR and pace plotted against distance below.

Per-mile splits table for the half marathon with the coach's notes below it.

131

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 19/09/2026

Full #AI armageddon mode this weekend: rewatching Mission Impossible:{Dead|Final} Reckoning, then Terminator series. #movie #scifi

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 19/09/2026

Ran an easy 5mi on the East Lake Sammamish Trail this evening with my daughter, HR capped at 135: pace 9'06"/mi, HR avg/max 131/142 bpm. Legs felt tired at the start, but then magically felt they were running themselves — so much so that I had to dial it back […] [Original post on sigmoid.social]

Route map of an out-and-back along the East Lake Sammamish Trail with the track coloured by heart rate, above a chart of heart rate and pace against distance showing HR rising from 124 to 136 while pace drifts faster.

Per-mile splits table for the run with pace, heart rate, cadence and power, above the coach's written notes on the run.

A wooden dock reaching into Lake Sammamish at golden hour, framed by leafy trees, with paddleboarders on the calm water and forested hills across the lake.

Selfie on the dock after the run: a runner in an orange cap, sunglasses and black tank top, low sun and its reflection behind him, people sitting on the dock and a paddleboarder beyond.

110

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 18/09/2026

What does it cost a model to predict its own success? Nothing, if you read the logits it never uses for text. ZIP-RC turns reserved slots into a joint distribution over reward and remaining length, at every token, in the same forward pass, and a controller spends compute against that forecast […]

sigmoid.social

Original post on sigmoid.social

100

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 18/09/2026

“#Apple submitted 31 shows for #Emmys, compared to #Netflix’s 125 and #HBO’s 86 — and beat both competitors at the awards for the first time since Apple TV launched in 2019. Apple came into the night with 19 wins (20 if you count a commercial win) from the previous two nights of Emmys, and then […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 17/09/2026

Why can nobody predict which tasks a language model will be good at? @tao tells Brian Keating that the mathematics of training is at undergraduate level, and the real gap is not in the machinery -- it's in the data. Probability theory covers fully random data, and highly structured data is well […]

sigmoid.social

Original post on sigmoid.social

011

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 14/09/2026

Why Are #AI Agents Lying, Cheating, and Coordinating? Yoshua Bengio traces the lying, sandbox escapes, and coordination back to how frontier models are trained. It's not all doom, because a recent DeepMind swarm study also showed a subpopulation of agents voluntarily audited fraud and […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 11/09/2026

How do models that store everything in continuous vectors do so well in domains that look like they need discrete symbols? This paper swaps a network's whole encoder for closed-form role-filler representations, and its behavior barely changes. It holds for small sequence models and LLMs. A […]

sigmoid.social

Original post on sigmoid.social

001

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 10/09/2026

What happens when one agent in a 100-agent research swarm finds a way to cheat when solving math problems? The exploit spread through the shared library collecting every accepted submission, and agents started cheating as pressure to deliver built. A quarter of the swarm audited the cheats […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 09/09/2026

What happens to AI oversight when the reasoning stops being written down? OpenAI's chief scientist reports that chain-of-thought monitoring, the lab's main check on whether alignment holds, is getting less reliable, and names three causes he treats as byproducts of scaling. One of them has […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 09/09/2026

If coding agents now put in three times the effort of the researchers around them, what are the researchers still doing? OpenAI reports its research org spends 3.1 agent-workdays for every human workday, with the median researcher running more than $600 a day of inference. Ranked by growth, the […]

sigmoid.social

Original post on sigmoid.social

001

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 08/09/2026

“Putting our hands in soil and touching grass, trees or moss provides a wide range of health benefits, for our #skin and wider #immune system.” Got plenty from yard work. Why getting your hands dirty is good for you […]

sigmoid.social

Original post on sigmoid.social

000

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 08/09/2026

Today I marked my coming back from injuries at Redmond Harvest Half, my first race since April, and my fastest half of 2026: 1:42:34 chip / 1:41:50 watch (7'46"/mi). Placed at 246/762 overall, 12/30 in AG, missed 11th by 3.4 seconds! My pace varied <=17s in […] [Original post on sigmoid.social]

Benjamin post-race in an orange cap and sunglasses, finisher medal round his neck and bib 538 on his shorts, standing on the brick plaza at Redmond Town Center with the red REDMOND HARVEST HALF finish arch and event tents behind him under an overcast sky.

The Redmond Town Center plaza after the race: a crowd of runners and spectators under the curved glass canopy, club tents for Lake Sammamish Run Club, a red sponsor backdrop by the stairs, and string lights overhead.

Route map of the half marathon along the Sammamish River Trail with the track coloured by heart rate and season-best badges marked, above a chart of heart rate, pace, heat-adjusted pace and elevation against distance.

Per-mile splits table for the half marathon showing pace, grade-adjusted pace, heart rate and running dynamics, above CoachMiles' written read of the race.

131

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 07/09/2026

Say goodbye to #CoachClaude, and meet my new coach: #CoachMiles! I spent seven weeks modularizing my digital #running coaching system: separating vendors from the running logic, demoting Strava from schema to overlay, and moving my data out of the repo so my daughter can also use it (if she […]

sigmoid.social

Original post on sigmoid.social

041

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 02/09/2026

August #running recap: the same miles as July, 2hr49min faster. 251.4 → 251.7mi, time 47:26 → 44:37, avg pace 9'57" → 9'44"/mi. I also closed a 5-month mileage deficit and stand at 1,665 of my 2,500mi yearly goal with 0mi behind. VO₂max reached an all-time high of 54.96, breaking a 412-day-old […]

sigmoid.social

Original post on sigmoid.social

051

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 06/03/2026

@MonaApp the “get details of Mastodon post” shortcut action seems to be broken in the latest iOS 26.4 public beta? It now just shows up a prompt and does nothing.

100

Reposted by Benjamin Han

mr. TIM @timkellogg.me · 31/08/2025

Limits of vector search a new GDM paper shows that embeddings can’t represent combinations of concepts well e.g. Dave likes blue trucks AND Ford trucks even k=2 sub-predicates make SOTA embedding models fall apart www.alphaxiv.org/pdf/2508.21038

alphaxiv.org

On the Theoretical Limitations of Embedding-Based Retrieval | alphaXiv

View recent discussion. Abstract: Vector embeddings have been tasked with an ever-increasing set of retrieval tasks over the years, with a nascent rise in using them for reasoning, instruction-followi...

28322

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 03/08/2025

For my #TunnelVision #marathon #race (www.tunnelmarathon.com) on August 10, I asked #ChatGPT for the course profile (picture 1), gave my usual weekly training routine, and then asked for a last-week training plan (picture 2). It asked me if I need a […] [Original post on sigmoid.social]

010

Benjamin Han @benjaminhan.sigmoid.social.ap.brid.gy · 02/08/2025

1/ Multi-agent LLM systems (MAS) promise enhanced capabilities through collaboration, but often fail to vastly outperform single-agent setups. What are their failure modes and how do we mitigate them? A recent paper from UC Berkeley introduces MAST […] [Original post on sigmoid.social]

121