Analyst, writer, systems thinker, forecaster. Formerly nightlife and crisis comms for Fortune 50s. AI, markets, culture, and the future etc. I’m @mgogel hi!
New York, NY
Joined February 2009
- Tweets136K
- Following827
- Followers8.7K
- Likes67.2K
I don’t know but ASD-STE100 was so two weeks ago.
Andrej Karpathy@karpathyUnited States
United StatesConnected via United States App StoreAccount-level information, not a live location or per-post device.
We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:
Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better:
Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better:
Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better:
Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work!
In summary:
- As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding.
- Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.
Interesting.
“The perfect fog of war” - October 1, 2026
A study of AI agents, the shift from skills to plugins, autonomous finance, crypto and tokenization, military realignment, and the changing architecture of attention.
The old conception of the fog of war depended on deprivation. A system does not need to make something invisible if it can make everything surrounding it irresistible. The fog of war is no longer necessarily the absence of information. It is the perfection of attention.
I have been marking versions of this condition for years, usually before I understood what I was marking. The crossing of the swords, anthracite, the beast and the recurring nine-day intervals belong to that record because they preserve an earlier state of observation before whatever followed could reorganize the memory of what came before it. Some later appeared significant and some led nowhere. Both belong in the archive, because a pattern that survives only after its failures have been removed is not a pattern. It is editing.
April 2026 matters because much of what surrounded the experience can now be separated from recollection. On April 8, after thirty-eight days of war and the beginning of a fragile ceasefire with Iran, senior American military officials said U.S. forces remained positioned to resume combat if diplomacy failed, with more than fifty thousand American troops still in the region. On April 13, after negotiations failed to produce a settlement, the United States began a naval blockade of Iranian ports. Iran temporarily reopened the Strait of Hormuz on April 17 while warning that it could close the passage again, and on April 22 Iranian forces seized two commercial ships attempting to leave the Gulf. The word ceasefire therefore described a period that also contained military readiness, blockade, disrupted maritime movement, negotiation and the possibility of renewed combat.
At almost exactly the same time, another field was becoming unusually easy to follow. Coachella occupied the cultural foreground across April 10 through 12 and April 17 through 19. On April 10, OpenAI was publicly teaching skills as reusable workflows, packages of instructions and supporting material that allowed a method to persist instead of being rebuilt prompt by prompt. During the second half of the month, skills repositories, coding agents, context systems and increasingly autonomous workflows spread rapidly across the technological foreground, while Karpathy’s late-April discussion of the movement from vibe coding toward agentic engineering supplied a name for a change already underway.
That technological wave was not something my feed invented. It existed outside me, but the feed made it feel nearly complete.
The asymmetry is what stayed with me. The geopolitical field required constant reconstruction because the significance of each development depended upon what happened next. The technological field unfolded with far less resistance. A discussion of reusable skills opened into coding agents, which opened into memory, context, verification, delegation and autonomy, and each of those led naturally into another subject I already wanted to understand. The material held attention because it was consequential enough to justify the next hour.
Apparently, I am spam.
I’ve been detected by X’s automated systems as potentially containing spam, so my posts are hidden from recommendations to non-followers, since, well, long before this feature came live, though there’s no historicity to it.
I took a look under the hood, where you can check your visibility and shadow-ban status:
Anyway, I always like confirmations. x.com/i/under_the_hood
A prompt share and a little self-knowing.
Sooooo I wanted to know something. How much do I actually read? It’s also a kind of AI screen-time measure.
I’m reading roughly one book’s worth of AI-generated text every week, about 50–70 standard books a year, requiring around 5–7 hours of reading in a typical week.
That works out to roughly 72,000–104,000 words of AI output every week, about 10,000–15,000 words per day, or 37–54 book pages per day at 275 words per page.
So I wondered, if I examined my entire AI chat history from my very first session until now, and assumed I actually read 100% of every response the models gave me, how much am I really reading every week?
Not tokens. Actual words, pages, books, and hours of reading.
I am turning AI usage from an app-engagement metric into an information-consumption metric. “Hours spent with AI” and “words actually read” describe two very different things.
So I wrote this prompt:
AUDIT full accessible AI-chat history, first session → now.
ASSUME read_rate = 100% of assistant/model outputs.
MEASURE wherever possible. Estimate only where necessary.
RETURN:
• total_output_words
• avg_words/week
• avg_read_time/week @ {200, 250, 300} wpm
• avg_books/week @ 80k words/book
• avg_pages/week @ 275 words/page
• lifetime {words, hours, books, pages}
• rolling {30d, 90d, 365d} vs lifetime baseline
• accessible_history_% + confidence_range
Do not infer whole-history volume from a small sample when broader records exist.
Label every result:
{MEASURED | ESTIMATED | UNKNOWN}
FINAL:
“Average weekly AI reading load ≈ X–Y hours, equivalent to A–B standard books.”
I think this is a much more interesting way to measure AI use than messages sent or hours spent in an app.
How much have you actually read?
ALT A prompt share and a little self-knowing. Sooooo I wanted to know something. How much do I actually read? It’s also a kind of AI screen-time measure. I’m reading roughly one book’s worth of AI-generated text every week, about 50–70 standard books a year, requiring around 5–7 hours of reading in a typical week. That works out to roughly 72,000–104,000 words of AI output every week, about 10,000–15,000 words per day, or 37–54 book pages per day at 275 words per page. So I wondered, if I examined my entire AI chat history from my very first session until now, and assumed I actually read 100% of every response the models gave me, how much am I really reading every week? Not tokens. Actual words, pages, books, and hours of reading. I am turning AI usage from an app-engagement metric into an information-consumption metric. “Hours spent with AI” and “words actually read” describe two very different things. So I wrote this prompt:
The next frontier is emerging as adversarial entities move toward “Neuromorphic Strike Systems,” AI architectures that map the collective neuro-semantic network of a population to identify minimum-effective-dose stimuli. Instead of spamming disinformation, these systems identify the precise sequence and emotional valence of inputs most likely to trigger cascading societal phase transitions, including mass panic and institutional paralysis, through structural resonance, bypassing rational evaluation almost entirely.
Yes I’m still early and you will not be able to properly verify this.
MG
ALT The next frontier is emerging as adversarial entities move toward “Neuromorphic Strike Systems,” AI architectures that map the collective neuro-semantic network of a population to identify minimum-effective-dose stimuli. Instead of spamming disinformation, these systems identify the precise sequence and emotional valence of inputs most likely to trigger cascading societal phase transitions, including mass panic and institutional paralysis, through structural resonance, bypassing rational evaluation almost entirely. Yes I’m still early and you will not be able to properly verify this. MG
Yo so I knew I remembered that number, and I knew I’d posted about it before.
Seeing the news now around A7/A7A5, Russia, Iran and the shadow-banking network on this latest “cycle,” I went back and checked.
There’s an interesting story in the timestamps.
On January 15, 2023, I wrote about Russia eventually building crypto infrastructure that other governments could not simply shut down, and the strategic implications if it became large enough.
A7A5 did not exist yet.
What did exist was the broader Russian push toward sovereign digital-payment infrastructure. The Bank of Russia had launched testing of its digital-ruble platform in February 2022 and had already completed the first digital-ruble transfers between individuals. The pilot involving real digital rubles would begin in August 2023.
Then, by December 5, 2025, I was writing specifically about A7A5, the ruble-backed stablecoin, describing it as cross-border payments infrastructure being used to evade or mitigate Western sanctions.
Then came October 1, 2026.
The U.S. Treasury took a much broader action against the A7 Network. OFAC sanctioned the A7 Network as a significant transnational criminal organization. FinCEN proposed a rule that would prohibit covered U.S. financial institutions from transmitting funds involving A7 Network Sub-Agents, and FinCEN issued an alert giving financial institutions indicators for detecting and reporting activity associated with the network.
Treasury now describes A7 as a Russia-linked shadow-banking network used by Iran to evade sanctions. It says the network’s Sub-Agents processed more than $17 billion between January 2025 and June 2026, while A7 itself claimed more than 7.5 trillion rubles, roughly $91.5 billion, in total transaction volume as of January 2026. Treasury explicitly identifies A7A5 as the blocked, ruble-backed token issued by sanctioned Old Vector LLC and says the A7 Network created it to enable its members to evade sanctions and transact internationally.
So I wasn’t imagining having seen the number before.
The chronology is the interesting part. Russia’s digital-ruble infrastructure was already entering testing in 2022. I wrote the broader Russia-crypto thesis in January 2023, before A7A5 existed. By December 2025, I was writing about A7A5 specifically. Then, on October 1, 2026, Treasury formally moved against the broader A7 financial architecture.
300 days separate my A7A5 post from this Treasury action.
Noting the timing. And remembering.
Today, Treasury took unprecedented action against the A7 Network, a shadow banking network with ties to Russia used by the Iranian regime to evade sanctions as part of Operation Economic Outcast.
Treasury’s @FinCENnews proposed a rule that would prohibit transmittals of funds regarding transactions involving the A7 Network’s Sub-Agents.
Additionally, FinCEN issued an Alert to help financial institutions detect and report suspicious activity related to the A7 Network.
Treasury’s Office of Foreign Assets Control sanctioned the A7 Network as a significant transnational criminal organization.
home.treasury.gov/news/press…
Show every time between 2016 & 2026 when I independently expressed concern about the world, reconstruct what information existed that day, show the S&P 500/VIX/BTC/oil response over the following 1, 7, 30 and 90 days, and tell me whether the idea preceded or followed the publicly visible signal for @mgogel
ALT Show every time between 2016 & 2026 when I independently expressed concern about the world, reconstruct what information existed that day, show the S&P 500/VIX/BTC/oil response over the following 1, 7, 30 and 90 days, and tell me whether the idea preceded or followed the publicly visible signal for @mgogel
