Probabilities 🪨
Georgia
Joined November 2014
- Tweets285
- Following125
- Followers106
- Likes153
Looking to HSL (Humanity's Last Exam), Claude Mythos/Fable 5 scored ~64% with tools enabled (According to Anthropic). This puts it under the >85% needed to be called super intelligent. But well within 56% – 85% for a "Human Expert Level" classification. For Comparison...
1
76
If this style of policy holds/grows in the future, it creates a stronger incentive for developing models outside current top labs. Even if a step behind the flagship ones, it creates a space in the market where they could be more competitive than the available flagship options...
1
22
Going after general intelligence with foundational models in the current state feels so much like brute-forcing encryption. It can achieve the objective in most cases, but is it the most eloquent/efficient solution?
17
Nikoloz retweeted
A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they end up in the general region you want them in.
As I was going through "A Brief History of Intelligence", it has been fascinating to be reminded that the debate on whether LLMs can or cannot scale to AGI, has been truly in the works for at least 35 years...
@maxsbennett
1
33
The time it took evolution to develop complex steering via Nematodes versus the ability to develop language in Humans is a lifetime versus a blink. I thought that was interesting, and also how it ties back to one of my favorites, Zima Blue from LDR.
1
32
Been using Claude Code for a few days now.
Last week I built smth: a personal assistant that pulls my Gmail, Calendar, WhatsApp and LinkedIn every morning and sends a structured daily briefing to my inbox.
Runs completely locally. $0 extra on top.
1) Here's how it works 🧵
ALT Flow chart for Claude Code local personal assistant.
1
2
55
2) The part I spent most time on was the structure itself. I didn't want a data dump, I wanted it to think about what actually matters.
It opens with a quote picked by the AI based on what's in your schedule that day. Not generic.
Then 6 sections →
ALT Daily Briefing Structure
1
18
This is the actual email snippet it sent me.
Runs fully offline; none of your emails or messages touch an external AI server. No expensive LLM API add-ons.
@ollama_ai @AnthropicAI
If you're curious how any part works, drop a reply 👇
ALT Email Snippet from the Daily Brief
14
Nikoloz retweeted
just another quick update on this research paper from *checks watch* 2 whole weeks ago: as it turns out, the new opus 4.6 data point is so far out of distribution that using the *same* methods from their paper to get a sigmoid fit results in a asymptote 2x lower than reality
UPDATE: here is the version a lot of people asked for. We fit our model of sigmoid(base capabilities) * sigmoid(reasoning) as well as the exponential curve on all models except the "unseen 3" (Gemini 3 pro, Claude Opus 4.5, and GPT 5.2) on the v1.1 dataset. This is an apples-to-apples comparison, and we believe both models are still plausible hypotheses. I hesitate to make any claims from 3 OOS points.
Part of the issue is that there were very few reasoning models in the original dataset, and so the inflection point of the sigmoid(reasoning) term is noisy.
Again, our point is to offer a competing hypothesis to raw exponential growth -- what we may be seeing is stacked S-curves of key innovations. If this is true, progress may begin to plateau without new innovations. All this is stated clearly in our paper, but we will make it clearer in an update.
Just wrapped up On the Edge by @NateSilver538:
1) Khachapuri? Didn’t see that coming! 😆🧀
2) A great window into the River, risk, philosophy, and thinking frameworks for high-impact tech developments 👏
34
Nikoloz retweeted
For the tenth consecutive day in Tbilisi, citizens are protesting against Georgian Dream's pro-Russian policies.
Today, there are no special forces or police on the streets. Instead, groups of violent men dressed in black are running around, beating people.
#GeorgiaProtests
An interesting read about language modernization and its frequent drawbacks by George Orwell
orwellfoundation.com/the-orw…
1
29
How to access GPT 4 without paying monthly subscription you ask?
Well @bing chat now runs on it!
@Microsoft - the benefits of being the largest "shareholder" in Open Ai – at Georgia
1
101
In 0 to 1 Peter Thiel really goes after the idea of monopoly and skipping competitive markets. Dilemma in a tech world is if it really is possible with a lurking danger of buyouts or emulation. A recent hopeful example of sort of addressing this is Open Ai.
Although Microsoft's – at Georgia
1
65
Multiple billionaires backing it. So the lesson here is if giants are too big and powerful learn to collaborate, so at least only profits are shared and technological power concentration is not aided. Is a startup capable of standing alone in lean-competitive market of giants? – at Georgia
1
38
That remains to be seen.
You can read about super interesting Open Ai structure by @gennarocuofano here fourweekmba.com/who-owns-ope… – at Georgia
