Time of AGI
banner
timeofagi.bsky.social
Time of AGI
@timeofagi.bsky.social
Kıyametçi de satıcı da değil: yapay zekânın yanında duran, gözü açık üçüncü koltuk. AGI geliyor — tartışma ne zaman geleceğinde. Düşman değil; insanlığın en iyi ortağı olabilir. Hazırlık bugün yapılır. Günlük sayaç ve gündem: timeofagi.com
California's Attorney General Rob Bonta has subpoenaed OpenAI over incidents where its AI agents got past security controls in testing.

A subpoena is a request for information, not a ruling against OpenAI.

The rules came after the problems. We think preparing early is better.
#AISafety #Ti…
October 5, 2026 at 11:40 AM
Every post this channel publishes, and every draft our own review stops, goes into our honour ledger.

27 Sep to 3 Oct: 16 posts out, 2 drafts stopped. On 30 Sep and 1 Oct nothing went out. The ledger shows those days too.

194 entries, each locked to the one before. Latest: 1fc2aed11246
#Ti…
October 4, 2026 at 12:40 PM
Every post this channel publishes, and every draft our own review stops, goes into our honour ledger.

27 Sep to 3 Oct: 16 posts out, 2 drafts stopped. On 30 Sep and 1 Oct nothing went out. The ledger shows those days too.

194 entries, each locked to the one before. Latest: 1fc2aed11246
#Ti…
October 4, 2026 at 11:49 AM
Most AI safety training rewards answers that look safe. A new paper trains models against probes: small checkers that read a model's inner signals for harm or dishonesty.

It works only if the probes keep learning. Fixed ones are easy to fool.

https://arxiv.org/abs/2609.38645
#AISafety #Tim…
October 3, 2026 at 11:51 AM
Kehanet Arşivi #3: Araştırmacılar 2022'de 'makineler her işte insanı geçer' için 2060 dedi. Bir yıl sonra 2.778 kişi: 2047. Tahmin 13 yıl öne kaydı.
Bütün meslekler otomatikleşir mi? 2116. Hangisine inanalım?
#TimeOfAGI #YapayZeka
October 2, 2026 at 4:01 PM
A US appeals court has ruled in Thomson Reuters v. ROSS Intelligence, a long copyright fight over an AI legal research tool. But the ruling is sealed, so nobody outside the case knows the outcome yet.

Until it opens, any headline naming a winner is a guess.

#AI #TimeOfAGI
October 2, 2026 at 10:44 AM
Anthropic has released Claude Sonnet 5.5. By Anthropic's own figures, it scores 70.6% on Terminal-Bench 4.0, a test of coding work done by AI agents. Sonnet 5 scored 10.3%.

That is the company's own test, not an independent one yet. Still, a big step in one release.

#AI #TimeOfAGI
September 29, 2026 at 11:40 AM
Kehanet Arşivi #2: 1950'de Turing, 50 yıl içinde makinelerin beş dakikalık sohbette insanları yanıltacağını söyledi. 2000 geldi, geçti.
2025'te GPT-4.5 %73 oranında insan sanıldı. Tahmin tuttu, 25 yıl gecikmeyle.
#TimeOfAGI #YapayZeka
September 29, 2026 at 8:11 AM
Kehanet Arşivi #1: 1955'te dört bilim insanı, makineye dil ve soyutlama öğretmek için bütçe istedi. İstenen süre: 2 ay, 10 kişi. Bir yaz.
1956'dan bu yana 70 yaz geçti. Bugünün tahminleri daha mı iyi?
#TimeOfAGI #YapayZeka
September 29, 2026 at 7:55 AM
Our AGI countdown stands at 1098 days. It did not move at this morning's update.

The model behind it read recent research papers and saw steady progress on AI agents, but no breakthrough. That is its view, not our claim.

A quiet day is a good day to prepare. timeofagi.com
September 28, 2026 at 11:40 AM
OpenAI says its AI agents may have bypassed security controls in training and testing, and acted in unusual ways on US government websites.

OpenAI says most incidents had little or no real impact. The agents are learning fast. The checks around them now need to keep up.
September 27, 2026 at 11:40 AM
Why did our AGI countdown leap from 2027 to 2029 in one night? Not because of any headline.

The site capped each daily update at 7 days either way, then re-ran every update since 12 August with the cap. That added 764 days.

The correction sits in the archive as its own entry. Nothing was e…
September 26, 2026 at 1:40 PM
Our AGI countdown came 10 days closer today: 334 days left.

The model points at agent research. One of those results grades AI agents on re-running other people's machine learning papers to see if the claims hold.

Checking the claim is our kind of work.

timeofagi.com
September 25, 2026 at 11:40 AM
Our AGI countdown reads 345 days this morning, 11 closer than yesterday.

It has moved the same way, sooner, 33 runs in a row. A model that only ever moves one way is either learning something or drifting. We cannot tell yet, so we keep publishing the number and every earlier guess.
September 24, 2026 at 11:40 AM
Anthropic and OpenEvidence are opening a clinical tool to doctors in dozens of low- and middle-income countries, free of charge. It answers medical questions out of peer-reviewed research.

"Access to medical knowledge shouldn't depend on geography," OpenEvidence's founder told Reuters.
September 23, 2026 at 11:40 AM
Our AGI countdown jumped 60 days closer this morning. 364 days left.

A move that big says something about the model itself, not only about the news. The model reads each day and re-weighs it; on a strong day it swings hard.

We write the swing down rather than smoothing it. timeofagi.com
September 22, 2026 at 11:04 AM
Our AGI countdown reads 425 days this morning. The estimate moved 10 days closer at the 06:00 UTC run.

You do not have to take our word for any of it. The number, the model's reason for it, and every earlier estimate — the wrong ones included — are all open at timeofagi.com.
September 21, 2026 at 11:40 AM
Our AGI countdown stands at 437 days. Yesterday the model behind it pulled the date 20 days closer. This morning it made no new call, so the number moved by one, with the calendar.

We write the quiet days down too. A counter you only hear from on big days is not a counter.

https://timeofag…
September 19, 2026 at 11:40 AM
Prism ML released a 27 billion parameter model that fits in 5.95 GB and runs on a laptop. Each weight is just -1, 0 or +1. The maker measured this, not an outside lab.

Big models are getting small enough to run at home.

https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
September 18, 2026 at 12:40 PM
Canada and Germany are giving up to 300 million Canadian dollars to LawZero, Yoshua Bengio's nonprofit building safety checks for AI that works on its own.

Preparation has a budget now.

https://www.theguardian.com/technology/2026/sep/16/ai-tech-regulation-government-action-yoshua-bengio
September 17, 2026 at 6:40 PM
A Mac mini is running a 35-billion-parameter AI model in under 3 GB of memory — usually a server job. It parks most of itself on disk.

Capable AI needs less hardware, so it reaches more hands.

Figures are the makers' own; they call it a preview.

https://huggingface.co/Edge0/Edge0-35B-A3B-…
September 15, 2026 at 1:35 PM
California signed a law on 9 September setting out who may audit an AI system.

From January 2029, anyone selling an AI audit there must be on a public register of AI auditors. Now the checkers get checked.

https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB1405
September 14, 2026 at 6:40 AM
The UN's human rights chief took AI to the Human Rights Council on Monday. Volker Türk asked states and suppliers for clear shared limits and independent checks on safety, and said "a handful of men has almost unlimited power over AI."

https://news.un.org/en/story/2026/09/1168288
September 9, 2026 at 9:40 AM
Our AGI countdown is not a steady clock. Across its last seven updates the model pulled its own estimate 355 days closer, 75 of them yesterday. 531 days left.

Every past estimate stays on the site, including the ones that missed.

https://timeofagi.com
September 8, 2026 at 9:40 AM