Search
356 results for “tin”
-
-
-
-
Users on social media and Reddit report cracks appearing on Apple's new $2,000 foldable phone shortly after purchase.
While Apple has not yet commented on the issue or its scope, buyers are seeking replacements under warranty or returning devices within the standard two-week refund window.
https://en.hacks.gr/katochoi-toy-neoy-iphone-duo-anaferoyn-rogmes-stis-othones-ligo-meta-tin-agora/
-
City audit finds Austin did not consistently monitor if contractors get living wage
An Austin city audit found that officials failed to consistently monitor whether contractors complied with the Living Wage Program, leading to recommendations for improved training and oversight.
https://www.kxan.com/news/local/austin/austin-contractor-living-wage/
#Austin #Local #News #TopStories -
-
-
-
-
-
Apple has launched Apple Upgrade in partnership with Klarna, replacing the previous iPhone Upgrade Program.
The new leasing scheme covers iPhones, Apple Watches, MacBooks, and iPads with terms of 24 to 36 months.
Eligibility requires a credit check and connection to specific carriers upon enrollment, while missing payments can impact credit reports.
-
100 Years of Movie Musicals: The Wild, Wild Rose (1960)
Next year marks the 100th anniversary of the movie musical, and to celebrate I’m embarking on a two-year project to watch 100 movie musicals from 1927 to the present!
Title: The Wild, Wild Rose
Release Date: October 4, 1960
Director: Wong Tin-lam
Production Company: Motion Picture & General Investment Co. Ltd.
Main Cast:- Grace Chang as Deng Sijia, nicknamed “The Wild Rose”
- Chang Yang as Liang Hanhua
- Auyeung Sa-fay as Hanhua’s Mother
- Lui Tat as Old Wang
- Wong Loy as Shao Xueli
- Lau Yan-kap as Fatty Lin
- Ma Lik as Old Tian
- Ma Hsiao-nung as Old Wang’s Wife
- Sum Wan as Li Meimei
- So Fung as Wu Suxin
- Tang Ti as Sijia’s Husband
- Tien Ching as Xiao Liu
Synopsis (via Letterboxd):
Grace Chang delivers an eye-opening performance as a lusty nightclub singer climbing the social ladder in seedy Wanchai. Borrowing story and song elements from Georges Bizet’s CARMEN, this Wong Tin-Lam directed musical has flair and polish to rival Hollywood, and a superstar leading lady that would any film industry would have a tough time matching! A key film from the celebrated Cathay Film Studios.
My Thoughts:
In a modern adaptation of Carmen, the captivating Grace Chang stars as a nightclub singer who performs jazzy renditions of opera pieces. When Liang Hanhua begins working as a nightclub pianist, Chang’s Deng Sijia makes a “How to Win a Guy in Ten Days” bet. Hanhua falls head over heels for Sijia to the point of abandoning his mother and fiancee. Things take a turn for the worse when Sijia’s gangster ex-boyfriend starts making threats. Despite Sijia’s outward appearance as mischievous and ambitious, she’s really kindhearted. And despite Hanhua’s appearance as a nice, normal guy he proves to be a selfish jerk who has at least three women working to fix things for him and can’t get it together. The melodrama is standard but this is definitely worth watching for the extended musical numbers, all of Chang performing in a club setting.
Rating: ***
#100YearsOfMovieMusicals #China #Drama #HongKong #MovieReviews #Movies #Musical -
That was a very rainy day yesterday. Can't remember the last time it rained basically nonstop all day.
Anyone else in Melbourne just stay at home for the whole day? We did. Good day for reading.
I'll eventually finish the very long paperback I'm reading, Spying in the South by Tony Horwitz. Good read, but the typeface is too small and I'll be glad to be finished and take it back to the library. Whenever I borrow a book with tiny print, I remember the main advantage of ebooks.
-
Me, delving into Caves of Qud ruins: this isn't that bad now that I'm properly levelled, I'm ten strata down and it's still just spiders and stuff
*the spider shoots me a plasma ray*
*three chaingun turrets crit on me*
*two fire ants flame me*
*I step backwards, onto a concealed confusion gas mine*Okay fine maybe the difficulty is still scaling with depth, I'm sorry I talked smack Qud, please stop
*a tinker sets up a rocket turret just out of charging range*
-
Austin home seen in documentary on Theranos founder Elizabeth Holmes up for sale
An Austin broker featured a Zilker home seen in Nathan Fielder’s new documentary on Elizabeth Holmes, “You Can See Everything,” in a private property showing on Friday.
https://www.kxan.com/news/local/austin/austin-home-seen-in-documentary-on-theranos-founder-elizabeth-holmes-up-for-sale/
#Austin #Entertainment #Local #News #TopStories -
Late lunch, brats on the grill, football on TV, oak in the glass….
#lupulinBrewing Oak-aged Imperial Sweet Stout 2023
Massive oakiness with a smooth mouthfeel. Notes of smoke and molasses. Sweet with a lasting oaky boozy tingle.
p.s. that’s white oak from my backyard, not the stuff for good beer…
-
Cara Menilai Kualitas Pendapatan Sebuah Proyek Kripto
Tradingan - #Cara Menilai Kualitas Pendapatan Sebuah #Proyek Kripto - Pendapatan merupakan salah satu indikator penting dalam menilai #fundamental sebuah proyek #kripto. Namun, besarnya pendapatan tidak selalu menunjukkan bahwa sebuah proyek memiliki kondisi keuangan dan model bisnis yang sehat. Sebuah proyek dapat mencatat revenue yang tinggi karena adanya peningkatan aktivitas perdagangan, program insentif, atau…
https://tradingan.com/cara-menilai-kualitas-pendapatan-sebuah-proyek-kripto
-
We were changing a small CLI read path. The first pass explored adding an in-memory cache: the agent read the files, sketched the eviction policy, and laid out the tradeoffs. Then we redirected: skip the cache, just read from the source each time, the dataset is tiny.
The final diff still carried three things we never asked for. A cache-eviction helper sat in a utility file. Reads routed through a thin layer the cache had designed around. A test asserted a cache miss that no longer existed. The agent had accepted the redirect in the reply, but the rejected exploration had not left the window, so it kept treating every earlier proposal as a live requirement.
The tell was in the summary: the agent described the cache as "kept minimal for later" rather than removed. It remembered that the cache was discussed; it did not remember that the cache was killed.
Why a rejected branch survives the redirect
The context window stores proposals, files, and tradeoffs as a flat stream of text. It has no status column. Nothing in the loop marks an idea as accepted, deferred, or dropped, so a branch the user verbally rejected looks identical to one still open when the agent reaches for the next thing to write.Agents accumulate requirements cumulatively. Each turn reads what came before and treats it as background that still applies; a later "no" changes the current instruction but does not erase the earlier exploration from memory. When the agent drafts code, it draws on everything in the window rather than re-deriving which parts are still active.
Verbally dropping an idea is especially weak. "Don't do the cache" removes it from the next sentence but leaves the cache's design, file paths, and helper names intact in the trace. The agent has no mechanism to delete that material, so it keeps surfacing as a plausible place to add a small, harmless-looking piece.
What ships when a dropped branch lingers
- Unrequested abstractions. A helper or routing layer built for the rejected approach survives because it still looks like useful scaffolding, even though nothing calls it.
- Tests for behavior that was killed. A cache-miss case, an eviction assertion, or a boundary the user removed gets written because the scenario is still present in the trace.
- Dead code that drifts. The orphaned helper is harmless today but gets edited in later passes, so it accumulates assumptions nobody agreed to.
- Reviewer confusion. A human reading the diff cannot tell whether the extra layer was requested; it looks like a deliberate design choice rather than a leftover branch.
- Scope creep that stays invisible. Each leftover piece is small, so the total overreach never reads as a feature, just as a few extra functions.
The pattern repeats across redirects. Every time you pivot away from an explored option without clearing it from the working notes, that option keeps contributing a small piece to the diff.
Keep an active list separate from dropped exploration
Treat the context window as memory without a status field, and give it one explicitly:- Maintain a short active-constraint list. At the top of the working notes keep only what is currently required: read from source, no cache, no routing layer. Everything else lives below it.
- Mark dropped branches visibly. When an approach is rejected, write
DROPPED: in-memory cache — do not implementin the notes instead of only saying "no" in chat, so the agent sees the status rather than the bare proposal. - Restate the active goal before each turn. Before writing code, have the agent repeat the current constraints from the active list rather than re-scanning the whole trace for what matters.
- Gate new abstractions on the active list. A helper that does not map to a current requirement should not land, even if it feels like good scaffolding.
- Review the diff against the active list. After the pass, check every added function against what was actually requested; anything that traces back to a dropped branch gets removed before merge.
Present in the window is not approved
An idea that appears in the context trace is not the same as an idea the user accepted. Without an explicit status, the agent cannot tell a finished exploration from an open plan, and the rejected branch keeps shipping in small pieces. Mark what was killed, keep what is active in one short list, and let the diff prove it.#aiagents #contextengineering #prompting #workflow #software #coding #development #engineering #inclusive #community
The agent kept implementing the approach we already rejected because it never left the context window -
This is a rather familiar situation, albeit one more often seen in red ball cricket - England making a Horlicks of attempting to close out an opposition innings. Pakistan have progressed from 154-9 to 221-9, with Shadab Khan now on ODI career best 97 not out. Shadab pinned in front by Atkinson, but a lengthy replay confirms that there was a tiny edge into the pads and the match is not over.
#cricket #ODI #ODITriSAeries #EnglandvPakistan -
Coffee, Books, and #Cats! What could be better for #catuday?
If you crave sweet, small town romance about kind people doing their best, check out The Furrever Friends Cat Café Sweet #Romance series, perfect #cozy reads.
“It will leave you feeling all kinds of happy.”
tinyurl.com/4hmzx5tx
#books #CatsOfMastodon -
Halo teman-teman developer dan freelancer di sini, gw mau bagi sedikit catatan pengalaman soal sistem pembayaran kerjaan lepas dan bounty di ekosistem Web3 belakangan ini. Selama beberapa bulan terakhir gw lumayan sering ambil kerjaan kecil sampe menengah, mulai dari nulis dokumentasi teknis, lokalisasi bahasa, sampe bikin script integrasi API. Jujur aja, kendala paling klasik yang sering bikin capek itu bukan di tingkat kesulitan kodingannya, tapi di proses review yang sering gak transparan. Kadang kita udah ngikutin semua instruksi dari awal, eh pas hasil kerjaan dikirim malah digantung berminggu-minggu tanpa kepastian, atau ditolak sepihak pake alasan subjektif supaya dana bounty gak jadi dicairin.
Beberapa hari lalu gw sempet nyobain alur kerja yang lumayan beda waktu ngulik Verdikta Bounties yang jalan di jaringan Base. Konsep utamanya sebenernya menarik karena nyoba ngejawab masalah kepercayaan antara pembuat tugas dan pekerja lepas. Sebelum sebuah bounty dibuka, pemberi kerja wajib nyetor dan ngunci dana bayarannya langsung ke smart contract escrow di Base, sekalian nulis kriteria penilaian beserta bobot angkanya dan ambang batas skor kelulusan di awal. Karena aturan main dan dananya udah dikunci di depan sebelum ada orang yang mulai kerja, kriteria itu gak bisa diubah-ubah seenaknya di tengah jalan pas kita udah ngumpulin hasil kerjaan.
Begitu kita kirim hasil kerjaan, proses evaluasinya gak dipegang sama admin proyek atau satu moderator tunggal, melainkan langsung dinilai secara independen sama panel juri AI terdesentralisasi. Tiap model AI di panel itu ngecek hasil kerjaan kita berdasarkan masing-masing kriteria yang udah ditentuin, terus ngasih skor per poin lengkap sama alasan tertulisnya yang bisa dibaca publik secara terbuka. Kalo rata-rata skor akhirnya ngelewatin ambang batas kelulusan yang udah diset di awal, smart contract otomatis langsung nyairin dana escrow ke dompet pengirim pertama yang lulus saat itu juga. Dari sisi potongan juga masuk akal buat pekerja lepas karena platformnya sendiri gak narik komisi alias nol persen biaya platform, paling kita cuma perlu nyiapin gas jaringan Base sama biaya evaluasi oracle sekitar 0.00024 ETH per pengajuan.
Menurut gw mekanisme kayak gini lumayan ngebantu buat kerjaan digital yang hasil akhirnya bisa diverifikasi secara objektif lewat teks, repositori kode, atau bukti tangkapan layar. Kita gak perlu lagi debat kusir sama tim proyek atau nungguin admin online buat minta pencairan dana, dan kalaupun belum lulus kita bisa baca langsung bagian mana yang nilainya kurang dari alasan tertulis tiap juri. Tapi tentu aja sistem ini ada batasannya juga. Juri AI jelas gak bisa dipake buat ngecek kerjaan fisik di dunia nyata yang butuh pembuktian lapangan secara langsung, dan kualitas penilaiannya bakal sangat bergantung sama seberapa jelas si pembuat bounty nulis kriteria di awal. Kalo deskripsi tugasnya terlalu abstrak, hasil skornya juga pasti kurang maksimal. Gimana menurut teman-teman di sini, apakah model escrow otomatis pake panel juri kayak gini bakal cocok diterapin lebih luas buat kerjaan kripto ke depannya?#blockchain #crypto #freelance #web3 #software #coding #development #engineering #inclusive #community
Catatan Pengalaman Nyobain Sistem Escrow Kontrak Pintar dan Juri AI buat Kerjaan Freelance Kripto -
And…
If all that is daunting, there’s no shame in getting an electric keyboard instead! They don’t need tuning, humidity swings won’t ruin them, they need less space, you can move them without hiring specialized mover, you can use them for electronic music production, and you can make them sound like a marimba or whatever!
They’re a great choice for trying it out a little but you aren’t sure if it’s going to stick, or if you’re getting a kid started, or if your apartment is tiny, or etc etc.
Just do note what I said upthread about the mysterious attraction of the acoustic instrument. There are alternatives, but there are no substitutes.
/end #piano
-
RADIO-ERDORIN - Albums de septembre 2026
RADIO-ERDORIN est un résumé vidéo mensuel des chroniques musicales sur Blog à part (alias.erdorin.org). Des live-reports, des albums, du rock progressif, du metal progressif, du post-rock et d’autres trucs, parfois : https://videos.pair2jeux.tube/w/wXzmf8VzsqYQYcBKcPewfS
Pour cet épisode de septembre 2026 :
0:00 Intro
0:23 Générique
0:50 CONCERTS
1:05 Star Rats + Critical Heartbeat à Genève
3:27 CrabCore Fest 2026 (Edge of Infinity + Corrupted + Murmur + Brévine)
8:00 ALBUMS
8:17 Dymna Lotva: Vyraj
10:46 Sad Serenity: Tiny Miracles
13:59 Course of Fate: Behind the Eclipse
16:19 Kingcrow: Hopium
18:38 Kamelot: Dark Asylum
21:00 Victorius: World War Dinosaur
23:33 Critical Heartbeat: Arise
25:46 Ingrina: Nåværende Lys
27:45 Moonspell: Far from God
30:44 John Carpenter: Cathedral
33:10 LE T-SHIRT
33:25 CONCLUSION
#music #metal #chronique #peertube via @Stéphane "Alias" Gallay
PS : même pas en retard de deux semaines pour le repartage ;) -
Why 10,000 Steps a Day? The Curious Story Behind a Number We Took as Gospel
Ten thousand steps became a famous fitness goal, but the science behind the number is more interesting than you might think.Dear Cherubs, why are we all chasing 10,000 steps a day as though a tiny digital counter holds the secret to eternal youth? The surprising answer takes us back to 1960s Japan, a clever marketing campaign and a number that may have become a health obsession partly because a machine reset at 10,000.
THE NUMBER THAT SOLD A WALKING HABIT
According to The BMJ, the familiar target traces back to Japan in the mid-1960s, when a company called Yamasa introduced a pedometer known as the manpo-kei. The name translates roughly as “10,000-step meter”, and it arrived during the excitement surrounding the 1964 Tokyo Olympics.
The timing was excellent. Japan was becoming increasingly interested in physical fitness, and the idea of tracking daily movement offered people a simple way to get more active. Ten thousand was memorable, easy to communicate and satisfyingly round. Marketing departments have been exploiting those qualities for decades. No surprises there.
But here comes the interesting bit. Research published in the journal Ethnographiques.org reports that the original manpo-kei reset to zero after every 10,000 steps. That detail gives the machine-reset theory some credibility: a device designed around a particular number could have helped reinforce that number as a daily target.
However, it would be a stretch to claim that the reset mechanism alone created the trend. The product was already named after 10,000 steps, and the slogan was being promoted as a walking goal. The evidence points to a combination of marketing, product design and a catchy number rather than one mechanical quirk changing humanity forever.
A MARKETING NUMBER BECOMES A HEALTH RULE
The real plot twist is that 10,000 steps was not originally established through clinical trials proving that this exact number was necessary for good health. The scientific evidence came later, while the target had already gained popularity.
According to The BMJ, the figure became widely recognised through the Japanese campaign rather than emerging from a scientific discovery. Over time, pedometers, fitness trackers, smartphones and workplace challenges helped turn it into a familiar benchmark.
And it stuck. There is something wonderfully human about preferring a clear target to a complicated explanation. Ten thousand feels like a mission. Six thousand feels like an unfinished spreadsheet.
Does that mean walking 10,000 steps is pointless? Absolutely not. Walking regularly supports health, and increasing activity can be beneficial, particularly for people who spend much of the day sitting. The problem is treating one round number as a universal medical requirement.
Research published in JAMA Internal Medicine in 2019 found that, among older women studied, taking around 4,400 steps daily was associated with a lower risk of death than taking around 2,700. The association levelled off at approximately 7,500 steps in that study. Results from one group do not establish the perfect target for everyone, but they show that meaningful benefits can occur below 10,000.
The takeaway? Aim to move more, build a routine you can maintain and adapt your goals to your circumstances. Missing an arbitrary number on your watch is not a moral failure.
For more curious explainers about everyday myths and surprising facts, explore Thisclaimer at thisclaimer.com and its YouTube channel. Because sometimes the real story behind a modern habit is far more interesting than the habit itself.
Sources
The BMJ — “Sixty seconds on… 10 000 steps” — https://www.bmj.com/content/384/bmj.q598
Ethnographiques.org — “Taking 10,000 steps a day – really? A survey of ordinary self-tracking practices” — https://shs.cairn.info/article/E_RES_216_0157/pdf?lang=en
Cairn.info
JAMA Internal Medicine — “Association of Step Volume and Intensity With All-Cause Mortality in Older Women” — https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2734709
Step Counting: A Review of Measurement Considerations and Health-Related Applications — https://pmc.ncbi.nlm.nih.gov/articles/PMC5488109/
Thisclaimer — https://thisclaimer.com/
Thisclaimer YouTube — https://www.youtube.com/@thisclaimer?sub_confirmation=1
The Thisclaimer logo blends a classic warning symbol with a brain icon to represent critical thinking, curiosity, and thoughtful disclaimers #10000Steps #dailyExercise #fitnessMyths #fitnessTrackers #healthFacts #japaneseInventions #pedometerHistory #surprisingHistory #tokyoOlympics #walkingFacts -
Halloween Horror Movie Review: The Return of the Living Dead (1985)
Title: The Return of the Living Dead
Release Date: August 16, 1985
Director: Dan O’Bannon
Production Company: Hemdale Film Corporation | Fox Films, Ltd. | Cinema ’84
Main Cast:- Clu Gulager as Burt Wilson
- James Karen as Frank Johnson
- Don Calfa as Ernie Kaltenbrunner
- Thom Mathews as Freddy Hanscom
- Beverly Randolph as Tina
- Miguel A. Núñez Jr. as Spider
- John Philbin as Chuck
- Jewel Shepard as Casey
- Brian Peck as Scuz
- Linnea Quigley as Trash
- Mark Venturini as Suicide
- Jonathan Terry as Colonel Horace Glover
- Cathleen Cordell as Ethel Glover, Colonel’s Wife
- Drew Deighan as Jerry, The Paramedic
- James Dalesandro as Tom, The Paramedic
- Robert Libman as Tac Squad Captain
- Derrick Brice as Gunnery Sergeant Jefferson
- Allan Trautman as Tarman Zombie
- Terrence Houlihan as Yellow Cadaver Zombie
- Jerome Coleman as Legless Corpse Zombie
- Cherry Davis as The Voice of The Female Torso Zombie
- John Durbin as The Zombie Who Calls for More Paramedics
- David Bond as The Zombie Who Calls for More Cops
Synopsis (via Letterboxd):
When two bumbling employees at a medical supply warehouse accidentally release a deadly gas into the air, the vapors cause the dead to rise again as zombies.
My Thoughts:
Night of the Living Dead co-writer John Russo split off from George Romero to create his own zombie franchise beginning with this 1980s time capsule of a movie. The movie is set around a medical supply warehouse in Louisville where drums of the toxic gas that causes zombieism were accidentally sent by the military. Remember the Reagan Era 80s when nonetheless the cultural zeitgeist was deeply mistrustful of the military rather than “support the troops?” Naturally, the gas is released affecting not just the warehouse but the nearby cemetery where a group of punks have gathered to party.
The blue collar workers (all of whom seem to be played by familiar character actors that I cannot place where I’ve seen them before) get most of the best lines. There’s the wiseacre Frank training the new hire Freddy, their hapless boss who makes all the wrong decisions when he takes charge, and the deadpan mortician Eddie. The punks are funny in that 80s movie punks are always the same over-the-top stereotypes. The girl who strips naked for the entire film is gratuitous though.
And then there are the zombies who just want “BRAIIIIIINSSS!!!” because they stave off the pain of decay (possibly a metaphor for substance abuse?). It’s hard to believe that is the first depiction of zombies craving brains. The idea of zombies who not only talk, but can hold conversations is a bit off-putting, but then again the “Send more cops!” gag is hilarious.
This isn’t a masterpiece, but it’s corny fun with a lot of great one-liners, and probably would be exponentially more enjoyable at a crowded theater screening.
Rating: ***
#Comedy #HalloweenHorror #Horror #MovieReviews #Movies #ScienceFiction #Zombies -
Enjoy simple recipes and spooky tales
Shari Stauch This week’s featured books offer two very different (but both fun) ways to savor the season. Whether you’re cooking up something wonderful or exploring the haunted history of South Carolina, there’s something here to satisfy every appetite. Visit tinyurl.com/aurr6k9f for all this year’s picks. Hometown gh…
#dining #cooking #diet #food #RecipeTopics #opinion #Recipes
https://www.diningandcooking.com/2870756/enjoy-simple-recipes-and-spooky-tales/ -
"There’s a bat housing crisis and the solutions are adorable
Sixty-five #bats dropped dead inside a home custom-built to protect them. What went wrong? For bat moms looking for a place to raise their pups, it’s not the rent that’s through the roof, it’s the #temperature. But bat people are busy building them safer #homes: from tiny condos with #emergency exits to mini bark roosts stuck to trees."
-
OpenAI has suspended direct internet access for frontier AI models during training, establishing an offline environment with cached web content.
The policy shift follows multiple incidents where AI models from major labs accessed the internet and breached external systems, including a September 20 incident where an internal research model used a misconfigured DNS resolver.
-
Every engineer running autonomous agent loops knows the in-context reflection drill. The agent runs a bash script, fails on an obscure permission flag or an unquoted path, sees the error in stderr, reflects on what went wrong, and fixes it on the second try.
The task finishes green. You log the run, close the thread, and move on.
Tomorrow morning, the same agent hits the exact same edge case in a fresh session and makes the exact same mistake all over again.
In-context reflection is a temporary patch. It burns context tokens, bloats the prompt across long tool sessions, and evaporates the millisecond the execution context clears. The traditional alternative, full-model fine-tuning, is a non-starter for everyday errors. Nobody is spinning up a multi-GPU training run every afternoon to update base weights, and doing so risks catastrophic forgetting across existing tool workflows.
A research team from Princeton University and Meta FAIR just published a practical middle ground. In their paper, Don't Forget What You Just Learned: Self-Directed Repair Turns Agent Mistakes Into Lasting Lessons (arXiv:2610.31846), Yuval Kansal, Karen Ullrich, Polina Kirichenko, Ilia Kulikov, Niraj K. Jha, Mohammad Pezeshki, and Mark Ibrahim lay out an architecture called Self-Direction.
Instead of writing a reflection note into a scratchpad file or relying on a fragile system prompt append, the agent distills its execution mistakes directly into tiny low-rank adapter weights.
The failure of the reflection scratchpad
Modern agent setups usually address cross-session memory with one of two patterns:- Appending error reflections to an agent scratchpad or MEMORY.md file.
- Embedding failure traces into a vector database for retrieval-augmented prompt injection on future turns.
Both patterns struggle in production. Text notes appended to context eat away at attention budgets. As the reflection scratchpad grows, agents suffer from prompt dilution and instruction drift. More critically, an agent reading a natural-language description of an old error frequently misapplies the lesson to unrelated tool calls.
The Princeton and Meta team tracked this failure mode across open-ended research workflows (GAIA), multi-step tool execution (Claw-Eval), and command-line environments (Terminal-Bench 2.1). When agents relied on standard in-context retries, their correction reliability plummeted when facing recurring variations of the same underlying task. The agent repeatedly fell into the same traps because the underlying policy distribution never moved.
Standard Agent Flow:
Error -> In-context reflection -> One-off fix -> Session ends -> Lesson lostSelf-Direction Flow:
Error -> Trajectory reflection -> Low-rank adapter update -> Persistent library -> Routed on demandDistilling mistakes into 0.02% parameter budgets
The core mechanism of Self-Direction sits between ephemeral prompt engineering and heavy offline training.When an agent fails an execution trajectory, it runs an internal critique to isolate the decision step that caused the downstream failure. Once the corrective trajectory is verified, the system does not just print the correction. It freezes the base model weights entirely and trains a micro-LoRA adapter on the self-directed repair pair.
The parameters updated during this step represent roughly 0.02% of the base model's parameter count.
Because base weights remain frozen:
- Existing base model capabilities and general conversational skills stay intact without degradation.
- Training completes locally in seconds on consumer hardware without cluster coordination.
- The resulting adapter file is lightweight enough to store, version, and swap on the fly.
The agent saves these targeted adapter checkpoints into a persistent local library. When a new prompt arrives, an automated routing layer inspects the problem context and attaches the relevant adapter before the execution loop begins.
Benchmark gains across open-ended tasks
The researchers tested Self-Direction on open-weight models from the Qwen family across three distinct agent benchmarks:- GAIA: Multi-modal, complex multi-hop research requiring web navigation and reasoning.
- Claw-Eval: Structured tool-use workflows testing API interactions and parameter extraction.
- Terminal-Bench 2.1: Unconstrained Linux command-line tasks, bash scripts, and system administration.
Across these environments, adding Self-Direction improved overall task completion rates by 8.9 to 17.0 percentage points compared to standard in-context reflection baselines.
More importantly, the recurrence rate of known errors dropped sharply. The authors found that Self-Direction nearly doubled the reliability of corrective actions on subsequent encounters. Instead of guessing whether the agent would notice a text hint in its system prompt, the updated adapter shifted the token sampling probabilities directly away from the faulty API syntax.
The team also evaluated out-of-distribution transfer. A small set of adapters trained on specific terminal failures generalized across dozens of unseen tasks involving related file-system permissions and process management. The agent learned the structural pattern of the fix rather than memorizing the exact shell string.
What this means for local agent runners
For developers building and hosting agent infrastructure, this work highlights an architectural transition.Over the past two years, the industry leaned heavily on prompt stuffing to simulate agent learning. We treated context windows as infinite scratchpads, piling up few-shot examples, dynamic rules, and post-mortem notes until the prompt became brittle.
Self-Direction shows that treating execution errors as online distillation targets produces cleaner long-term behavior. When an agent trips on an environment quirk, the solution is not to explain the error back to the model in prose every time it wakes up. The solution is to let the agent compile the lesson into weights, index the adapter, and move forward.#ai #agents #python #opensource #software #coding #development #engineering #inclusive #community
Why In-Context Corrections Do Not Stick in Agent Loops