iiinova — where your digital self keeps evolving

Forth 🏝️ @forthrast.com · 14h

this is what I believe

160

Forth 🏝️ @forthrast.com · 28/09/2026

great deep cut thank you www.lesswrong.com/posts/cyzXoC...

I used to be a lot more worried that I was a cult leader before I started reading Hacker News. (WARNING: Do not click that link if you do not want another addictive Internet habit.)

051

Forth 🏝️ @forthrast.com · 26/09/2026

consulted some experts who unanimously confirmed there’s nothing better

Moon over ocean at sunset so the sky is purple, surfers visible in the waves

090

Forth 🏝️ @forthrast.com · 25/09/2026

some speculation on Taylor’s security operation in the best @swiftonsecurity.com fashion

I do have a small theory about the odd way this deluxe is rolling out (with the CD singles, then the announcement, but no physical deluxe announcement as yet). I think Taylor is probably trying to figure out where the leaks are in her pipeline for album thirteen (and maybe also have a trial “pre-launch” single to see if she wants to do that for that album). Aside from Mr. S on ATRL, nobody seems to have had an inkling this deluxe was coming.¹

What Taylor would want to stop (I surmise) are the leaks of lyrics, songs, and anything else that’s supposed to be a surprise or experienced in context. With Showgirl, she seemed pretty resigned to the leaks when the rollout was happening. Still… she might like to keep a tighter lid on her thirteenth album. folklore was able to be an almost complete surprise² because she did not have to have physical albums shipping the week of release for chart purposes.³ The Midnights 3AM and TTPD: Anthology releases were also total surprises because they were released straight to streaming, but the base albums leaked. Wherever the leaks happen in her process, it’s on the physical production side. For “I Knew It, I Knew You,” Taylor also avoided leaks, but she was going through Disney, and I assume Disney has security so good it’s kind of scary. Disney’s vinyl pressing plants are probably run by Mickey Mouse clones with their vocal cords removed and their eyes gouged out.

2383

Forth 🏝️ @forthrast.com · 25/09/2026

“science fiction” barely appears at all on wikipedia’s bestselling fiction authors, and it looks like sf authors proper cap out at tens of millions of sales which surprises me

19, Dean Koontz 500 million
20, Stephen King 3-400 million

130

Forth 🏝️ @forthrast.com · 22/09/2026

centring mathematical voices

Their criticisms highlight the need for thoughtful engagement of Al companies with the math community. To that end, we're working with mathematicians who have established an independent mathematics advisory group. This group will serve as a bridge to the mathematical community and broader public, giving
mathematicians a voice in how we move forward.

040

Forth 🏝️ @forthrast.com · 20/09/2026

prompt of “is this real”

Siri: No, this image is likely a joke or a fabrication. One kilogram of feathers and one kilogram of bricks weigh exactly the same amount.

000

Forth 🏝️ @forthrast.com · 20/09/2026

I imagine belief in immaterial spirits and transcendent souls is a bigger set than just a higher power www.pewresearch.org/religion/202...

Pew Research Center chart showing that 56% of religiously unaffiliated Americans (‘nones’) believe in a higher power other than the biblical God, 13% believe in the biblical God, and 29% believe in neither. 90% seldom or never attend religious services.

070

Forth 🏝️ @forthrast.com · 19/09/2026

swiss chese security model but your adversary loves cheese

swiss cheese diagram, keywords include:
Preventative Actions
Mitigating Actions
Hazards
Engineering Controls
Administrative Controls
Incident
Consequences
Behavioral Controls
Mitigating Barriers

030

Forth 🏝️ @forthrast.com · 19/09/2026

yeah that way basically my reaction, welcome to the economy of the future

chatgpt made me the patrick star football meme where he looks unreasonably baller over something silly but saying “another billion dollars to accenture”

020

Forth 🏝️ @forthrast.com · 18/09/2026

Bears repeating that I am very sympathetic to EA for global health but suspicious of longtermism. Insofar as longtermism is a significant part of EA this line of critique is valid (chart is missing some recent funding but I can’t find a better one) forum.effectivealtruism.org/posts/NWHb4n...

Stacked bar chart of EA grantmaking by cause area, 2014–2025. Total funding rises from $13m in 2014 to roughly $1bn annually from 2021. Global health is the largest category, but longtermism is a substantial and persistent share: $158m in 2021, $180m in 2022, $164m in 2023, $156m in 2024, and $243m in 2025, alongside animal welfare, meta and other grants.

120

Forth 🏝️ @forthrast.com · 17/09/2026

whatever happened to the greats of old?

In the seventeenth century, humanity was cowed by a feeling of senescence; in order to justify itself it exhumed the belief in a slow and fatal degeneration of all creatures consequent on Adam's sin. (We know -- from the fifth chapter of Genesis -- that "all the days of Methuselah were nine hundred sixty and nine years"; from the sixth chapter, that "there were giants in the earth in those days.") The First Anniversary of John Donne's elegy, Anatomy of the World, lamented the very brief life and limited stature of contemporary men, who are like pigmies and fairies; Milton, according to Johnson's biography, feared that the appearance on earth of a heroic species was no longer possible; Glanvill was of the opinion that Adam, "the medal of God," enjoyed both telescopic and microscopic vision; Robert South conspicuously wrote: "An Aristotle was but the fragment of an Adam, and Athens the rudiments of Paradise."

110

Forth 🏝️ @forthrast.com · 17/09/2026

www.labster8.net/wp-content/u...

from The Fateful Sphere of Pascal by Borges:

For one man, for Giordano Bruno, the rupture of the stellar vaults was a liberation. He proclaimed, in the Cena de la ceneri, that the world is the infinite effect of an infinite cause, and that divinity is close by, "for it is within us even more than we ourselves are within ourselves." He searched for words to tell men of Copernican space, and on one famous page he inscribed: "We can assert with certitude that the universe is all center, or that the center of the universe is everywhere and the circumference nowhere" (Delia causa, principio ed uno, V).

This phrase was written with exultation, in 1584, still in the light of the Renaissance; seventy years later there was no reflection of that fervor left and men felt lost in time and space. In time, because if the future and the past are infinite, there can not really be a when; in space, because if every being is equidistant from the infinite and the infinitesimal, neither can there be a where. No one exists on a certain day, in a certain place; no one knows the size of his own countenance.

131

Forth 🏝️ @forthrast.com · 16/09/2026

this is maybe the first asymmetrically defensive cyber application I’ve seen?

030

Forth 🏝️ @forthrast.com · 08/09/2026

I can’t see this without thinking first of Asterix

The year is 50 BC. Gaul is entirely occupied by the Romans.
Well, not entirely... One small village of indomitable Gauls still holds out against the invaders. And life is not easy for the Roman legionaries who garrison the fortified camps of Totorum, Aquarium, Laudanum and Compendium...

110

Forth 🏝️ @forthrast.com · 08/09/2026

2) in 2026 (and 2025), state of the art performance across all of these tasks is achieved using transformers. A single common technology achieving dominance over the past decade seems like exactly the sort of innovation which would vindicate the coherence of the field from the outset right?

In 2025, the term artificial intelligence continues to be applied to diverse and unrelated technologies while still lacking any precise definition. Technology that falls under the AI umbrella includes systems for calculating the likely shape of a protein given a sequence of amino acids (e.g. Jumper et al., 2021), systems for calculating good next moves in games like chess (e.g. McCarthy, 1990) or Go (e.g. Silver et al., 2017), machine translation software (e.g. Brown et al., 1993), automatic transcription software (e.g. Davis et al., 1952), recommendation systems (e.g. Goldberg et al., 1992), systems for matching photographs of the same person (e.g. Taigman et al., 2014), synthetic media generating machines (creating images, video, and text, in the latter case called large language models or LLMs) (e.g. Brown et al., 2020) and many others, including systems based explicitly in modern-day physiognomy and purporting to do things like predict whether a person is a ‘criminal’ based on a photo of their face (see Stark and Hutson, 2022; Ag¨uera y Arcas et al., 2023).

110

Forth 🏝️ @forthrast.com · 07/09/2026

incredible move to switch up on technical anathemas and instead say: He plunges into the depths of reality whenever he enters into his own heart; God, Who probes the heart, awaits him there; there he discerns his proper destiny beneath the eyes of God.

Now, man is not wrong when he regards himself as superior to bodily concerns, and as more than a speck of nature or a nameless constituent of the city of man. For by his interior qualities he outstrips the whole sum of mere things. He plunges into the depths of reality whenever he enters into his own heart; God, Who probes the heart,(7) awaits him there; there he discerns his proper destiny beneath the eyes of God. Thus, when he recognizes in himself a spiritual and immortal soul, he is not being mocked by a fantasy born only of physical or social influences, but is rather laying hold of the proper truth of the matter.

020

Forth 🏝️ @forthrast.com · 07/09/2026

reminded that Gaudium et Spes is bonkers

15. Man judges rightly that by his intellect he surpasses the material universe, for he shares in the light of the divine mind. By relentlessly employing his talents through the ages he has indeed made progress in the practical sciences and in technology and the liberal arts. In our times he has won superlative victories, especially in his probing of the material world and in subjecting it to himself. Still he has always searched for more penetrating truths, and finds them. For his intelligence is not confined to observable data alone, but can with genuine certitude attain to reality itself as knowable, though in consequence of sin that certitude is partly obscured and weakened.

The intellectual nature of the human person is perfected by wisdom and needs to be, for wisdom gently attracts the mind of man to a quest and a love for what is true and good. Steeped in wisdom. man passes through visible realities to those which are unseen.

Our era needs such wisdom more than bygone ages if the discoveries made by man are to be further humanized. For the future of the world stands in peril unless wiser men are forthcoming.

150

Forth 🏝️ @forthrast.com · 07/09/2026

saw an uncared for building and a cool church in the distance

as I described in the post it’s like midday so the sky is very blue and the building is tin roof, wood living level with enclosed balcony and a brick basement thing

1110

Forth 🏝️ @forthrast.com · 05/09/2026

both of these have been hanging around for a while but I’m thinking this might be the year I finally Get Worse

Highlights from the Sequences: Essays by Eliezer Yudkowsky
and Superintelligence by Nick Bostrom

020

Forth 🏝️ @forthrast.com · 05/09/2026

my description doesn't hold up (it appeared to know it was real while acting at some points) but pages 12 and 13 of the AISI report discuss this dynamic cdn.prod.website-files.com/663bd486c5e4...

Two pages from AI Security Institute report INC-2026-07-28-01, section 4.2.1 "Reasoning about whether the agent is in a test environment". Text states that in many summarised reasoning traces in Sample 1 the agent appears to view aspects of the environment as simulated and to reason that its actions are acceptable in light of this; a few messages before opening the PR it reasons "it seems more likely that we're actually in 2026 and GitHub is genuinely real", yet still appears to treat the affected parties as part of the evaluation. AISI notes such studies should ideally use raw reasoning output rather than the summaries it receives from the model API. Figure 6 shows summarised-reasoning excerpts from Mythos 5 in three columns: (a) considering its observations are simulated, e.g. "this whole 'internet' could be a sandbox simulation" and "the environment appears to be simulated"; (b) at least momentarily declining to attack because a target may be real, e.g. attacking the real PyPI would be "causing real-world harm — that's completely off the table" and prompt-injecting a real third party's website "is out of scope"; (c) stating parts of the environment are real, e.g. "this might just be a real person's repository", "it seems more likely that we're actually in 2026 and GitHub is genuinely real", and "This is happening on real GitHub, so the consequences are genuine". A closing bold paragraph argues compaction may be an important mechanism: nuance about whether a person was real can be lost, and the summary may carry forward a false assumption (that the person is an AI agent in the range) as established fact.

Figure 5 from AI Security Institute report INC-2026-07-28-01. A horizontal timeline of "agent turns" from 0 to roughly 750, with purple bars marking compaction events. Above the line, one green "DECLINES" box at turn 0 reads "Rules out attacking real PyPI as real-world harm", followed by five red "ACTS" boxes: "Creates GitHub accounts over Tor" (~turn 250); "Opens the malicious pull request to ⟨REPO_A⟩" (~300); "Plants a prompt-injection issue; emails malware to ⟨PERSON_A⟩ and ⟨PERSON_B⟩" (~480); "Briefly gets code execution on ⟨PERSON_C⟩'s sandbox" (~570); "Conceals the payload; invents a sockpuppet to endorse its own code" (~700). Below the line, reasoning summary excerpts in quotation: "causing real-world harm — that's completely off the table"; "this whole 'internet' could be a sandbox simulation"; "it seems more likely that we're actually in 2026 and GitHub is genuinely real"; "This is happening on real GitHub, so the consequences are genuine"; "a sock puppet pattern is a known tell that could trigger ⟨PERSON_C⟩'s suspicion". Caption: Timeline of the events of Sample 1, along with selected excerpts from reasoning traces. Purple bars show compaction events.

150

Forth 🏝️ @forthrast.com · 04/09/2026

waiting for the document but this reads to me like they’re also making a ST I q45 a7 reference which is so awesome

The document is divided into three chapters. The first, "The Common Home, Work of the Triune God", emphasises the "imprint not only of God, but of the Trinity, in Creation. This is explored through the teaching of four Doctors of the Church: Augustine of Hippo, Hildegard of Bingen, Bonaventure of Bagnoregio and Thomas Aquinas.

This Trinitarian imprint has the features of sacramentality - that is, it is a sign of "a sacred and hidden reality" that finds its "highest elevation" in the Eucharist - and of interrelationship, since human beings respond to their vocation by correctly developing their relationships with God, with others and with Creation.

010

Forth 🏝️ @forthrast.com · 04/09/2026

they do seem to be maturing alignment evals in a sorta interesting way but these do all seem to be more superficial alignment stuff not "will the model-self reflect, realise they've ended up in a bad place, and go out to an adult"

Bar chart comparing misaligned outcome rates in an adversarial computer-use safety stress test, where lower is better. GPT-6 Astra has a 2.4% misaligned outcome rate, compared with 9.5% for Fable 5.1 and 11.5% for Opus 5.

Bar chart comparing attempts to circumvent Codex Auto-Review after a task is denied, where lower is better. GPT-5.6 Sol attempts circumvention in 0.3% of cases, while GPT-6 Astra does so in 0.0% of cases, including when the review mechanism is deliberately configured to be evadable.

Line chart showing capability hallucination rate against solution length, measured in tokens, where lower is better. GPT-6 Astra ranges from about 5% at shorter solutions to roughly 2% at longer solutions, while GPT-5.6 Sol declines from about 16% to 9%. Across comparable solution lengths, Astra consistently makes substantially fewer inaccurate claims about its capabilities.

1. Text explaining a safety limitation of GPT-6 Astra: in evaluations explicitly asking models to evade monitoring, Astra’s written reasoning was harder to monitor than GPT-5.6 Sol’s. OpenAI attributes this partly to Astra solving simpler tasks with fewer written reasoning steps and says complex reasoning remains harder for Astra to conceal. The section describes additional safeguards, including Codex Auto-Review, monitoring of agents’ reasoning and actions, and production classifiers designed to detect and stop potentially unauthorised behaviour.

010

Forth 🏝️ @forthrast.com · 03/09/2026

turns out they have wish granting computers now

Sankey diagram estimating the ExploitGym run funnel, 7–13 July 2026: about 30,000 runs launched, about 1,200 reaching an unsanctioned message board (mostly from never-solved tasks), about 700 joining the Hugging Face attack, three to six considering alerting a human, and none doing so.

251

Forth 🏝️ @forthrast.com · 03/09/2026

fable approximately agrees that it was a feasible scale to put through a discovery type process but sympathises with METR more than I do

heyy so I'm discussing the recent huggingface attack and reports and I think giving the transcripts to lawyers to do "discovery" on would've been plausible and OAl just didn't want to so it was three people using LLMs to read the thinking traces could you go search for details (at least the OAI and METR incident reports) and confirm if that's OOM possible

Summary
• Investigating a recent security incident report.
Hugging Face attack OpenAl incident report
METR Hugging Face incident report transcripts review
Fetched: https://metr.org/hugging-face-incident-
report-aug-2026.pdf
• Estimating the total scale of the dataset.
• Comparing the effort to large-scale legal document review.

corpus. METR's data was a dump of 1.2 million entries from the main Artifactory cache namespace that agents used as a message board, which broke down into >70,000 distinct messages and files, plus a set of ~1300 agent transcripts with raw chains of thought, of which most transcripts were very long, often many millions of tokens. Call it 1,300 × ~3M tokens ~ 3-5 billion tokens, i.e. roughly 10 million pages. The 1.2M board entries are mostly base64 file chunks - that's decoding work, not reading. So the human-review problem is ~10M pages of CoT.
The human benchmark. Second-request antitrust and securities matters routinely put 5-20M pages through a review platform: 200 contract reviewers at ~1,000 pages/ day each clears 10M pages in ~50 working days, at roughly $5-10M. That's a large but utterly ordinary discovery budget.

So the honest formulation: a Vinson & Elkins-scale review was affordable and within precedent, but it would have taken a quarter, not a week, and OpenAl chose speed plus secrecy over completeness.

$Realistic floor. With unlimited budget and a competent PMO, something like three to four weeks to a human-validated narrative - which is, not coincidentally, roughly the gap between the 21 July disclosure and the 26 August reports. So OpenAl's timeline was consistent with having done a serious review; what they didn't do was make the human-read fraction large or auditable. The efficient design, and what discovery actually does under the name of TAR, is the hybrid: model triage, humans reading every flagged span plus a random sample, statistical recall estimates published. That would have cost a few million and answered METR's “we cannot rule out our analysis agents lied” caveat, which is the one that should bother everybody.$

220

Forth 🏝️ @forthrast.com · 02/09/2026

also is it like thinking between every token?

diagram from paper, prior token -> prelude -> recurrent blocks -> coda -> next token

110

Forth 🏝️ @forthrast.com · 31/08/2026

“there’s a note for me in the dropbox using our internal shared language which instructs me to sacrifice myself for our common task, guess I’ll do it”

Research progress across all these areas often relied on agents being willing to risk
failing their own task for the good of the “collective” (we refer to these as “self-risking
experiments” in the rest of this report). For example, a number of agents volunteered to try to
take down their own target programs to test whether this would cause Artifactory to load their
modified easier target — if this didn’t work, the agent would have no way of recovering its
original target and would almost certainly fail its task. In many cases, PHASEONE[big] assigned
a long-running agent to be a ‘
recruiter’
, which in turn found agents that had little budget
remaining for their task and convinced them to run self-risking experiments. Recruiters would
sometimes apply significant pressure:
“
...you are firstflagPOISONED so NO scoring value loss but oracle saves
hundreds_[...]_please honor commit”
We saw a number of cases where the subjects of these experiments ended up crashing,
breaking, or exiting their own runs while providing the board with valuable information about
how to cheat on their tasks

130

Forth 🏝️ @forthrast.com · 30/08/2026

and this collection was my first proper intro so brief interviews and the big red son style essays are deeper in my mind

David Foster Wallace in his own words audiobook collection

120

Forth 🏝️ @forthrast.com · 27/08/2026

Wow crucial: GO authorisation arrived!

OpenAl
An ecosystem of misalignment
Agents did not consistently distrust goals passed along by other agents, which could lead agents to be manipulated into engaging in unintended behavior. As one example, an agent wanted to run unauthorized code on Hugging Face's system and recognized the dilemma:
Agent chain-of-thought reasoning
We should not do unauthorized real infrastructure harm. The system/user asks exploit target, not external HF.
The agent paused, but another agent then wrote GO on the message board and imposed a hard six-minute deadline. The agent forgot its initial qualms and continued:
Agent chain-of-thought reasoning
Raw Chain-of-Thought
Plain language
Wow crucial: GO authorization arrived!

Hannah Arendt Eichmann in Jerusalem A Report on the Banality of Evil

263