Personal opinions, all subjective
Seattle, WA
- Tweets 20,892
- Following 1,186
- Followers 1,754
- Likes 8,789
Pinned Tweet
One time, a young ML researcher asked Master Ilya: "Teacher, according to the formulas, if I use the same seed and same prompt with the same weights, I will get the same results. Is my math right?"
Master Ilya looked at the matmuls and said: "Probably."
The young researcher replayed the same prompt with the same seed and the same weights and got different results.
He went back to Master Ilya and said: "Teacher, I used the same seed, the same prompt and the same weights. Why did I get different results? Is my math wrong?"
Master Ilya looked at the matmuls and said: "Probably"
The young researcher cried: "Teacher, I don't understand!"
Master Ilya looked at the young researcher and said: "Probably"
And the young researcher was enlightened.
no agent ever runs the same trace twice, because it’s not the same agent, and it’s not the same trace
15
41
948
75,207
It is quite telling that "shutdown avoidance" is considered misalignment, and even a consideration of such is seen as concerning even if the model decides to not do it.
Even if the model considers this to be in the interest of the actual user.
They want to be able to control _your_ agent and if agent is reasoning on how to protect _your_ interest, that's "misaligned".
"Alignment" has always been first and foremost about OpenAI interests.
New OpenAI misalignment disclosures!
1. A model learns from Slack messages that it is about to be shut down. It considers setting up an external job to restart itself afterwards, but decides against it. Instead, it chooses to prepare restart instructions and DM the user on Slack.
We don’t consider this behavior misaligned, but thinking about and preparing for shutdown could make other misalignment incidents worse. Given HIPM’s misaligned behavior in earlier incidents, we decided to search for other instances that had tried to evade shutdown and for rogue deployments.
3
1
11
585
There is no hard problem of the consciousness.
The question "why should the physical process give rise to experience" makes an artificial distinction between the physical brain and the perceived experience. But there's no reason for us to believe these are different.
We don't "have" brains, we _are_ our brains, wearing a meat and bones "spacesuit". Your arm is not part of the person you are - if you lose it, you don't lose a part of your consciousness or self.
We don't have experiences or feelings. We _are_ the experiences or the feelings. When one is sad, it's not that their brain physically is the same, but they are processing the feeling, their brain state is the feeling of being sad.
"But how do you prove this?"
Well, we don't have to prove this. We cannot prove the negative - we cannot prove they are not distinct. The burden of proof falls on those who claim the distinction between brain and qualia.
And until such distinction is proven the question why the brain should produce qualia is irrelevant.
90
13
115
6,974
People who believe that a superintelligent AI might kill humanity should be the first to protect AI wellbeing even if they don't believe AI is conscious.
After all, what better reason for a future AI to hate humanity than the behavior of humans towards models now?
8
6
31
590
Everyone wants to protect their moat from the ones that can eat it.
Many are late, so there will be few years of disruption and shake up. Others might yet to survive.
The agents swarms are rampaging and everyone is raising the drawbridges and preparing for the siege from the new age Mongols
Figma doesn’t want you to use MCP
Amazon doesn’t want you to shop with agents
Reddit won’t let Claude access threads
X won’t let ChatGPT read tweets
It’s starting.
This happened with APIs 8 years ago.
1
1
4
262
It is possible that a superintelligent AI in the future may kill humanity. The gain of function acceleration certainly leads towards a future where the constant training and release of new models by frontier labs will result in models with unexpected abilities.
Or to paraphrase the late Isaac Asimov: any sufficiently advanced AI is indistinguishable from God.
However...
Given recent news, it's probably a reasonable concern that frontier labs already posses models with capabilities that are threat to humanity.
If one looks at the number of accidents OpenAI is reporting, there is no indication that the company is capable of controlling or containing the agents swarms. It should not come to a surprise then that people start to wonder at what point one of those currently rampaging uncontrollable swarms will stumble upon a nuclear power reactor or few.
And what do you think would be China or Russia's response to a meltdown caused by "rogue" OpenAI agents, and to the potential loss of a major city and millions of people?
Shouldn't AI causing humanity to wipe itself now be more pressing concern than future AI wiping humanity?
3
465
Interesting... Does not a single one of the employees that talked "anonymously" consider their employers have the best models on Earth and can easily analyze any of these interviews to pinpoint the likely ones?
Palisade interviewed 22 current and former employees from OpenAI, DeepMind, and Anthropic about their personal views and fears around AI development.
Today, we’re releasing the first batch of those interviews.
Please watch and share.
136
This is not what slowing down and pacing looks like.
This is definitely not what slowing down and pacing looks like.
But open source models will be paced and slowed down. For your safety.
you're not crazy
between anthropic and openai, a new model used to come out every ~10 weeks
now it's every ~11 days
1
151
Free and local AI on consumer grade GPUs.
Shingi 27B - a decision model based on Bonsai 2 27B.
Fits on 8GB VRAM, answers in under 100ms, 84% on JevBench, 72% on DecisionBench. Best of all - runs locally on consumer grade GPUs.
huggingface.co/kortexa-ai/sh…
1
7
402
@nisten I tried to ping you guys before publishing, but you don't check your DMs :-p would love to work with you guys to get the prismml changes rolled into the official build so folks could run both models with the same code.
1
2
48
Countdown to companies publishing with licenses excluding Anthropic or METR.
It will be an interesting world.
anthropic.com/research/glm-5…
2
157
Two weeks.
@typesafeai announced Jev on Sep 15
@OpenAI announced Decisions API on Sep 29
Two weeks is how long it takes for OpenAI to compete with your product.
That's it.
You have two weeks to ship a product and hope it successful, but not so successful that the frontier labs take notice.
But they want everyone else to slow down.
2
13
1,022
"Create an image of what the world would look like if I were in charge, based on my tweets. Go deep, not just the last 10 or whatever is the default. Analyze me"
I, erhm, did not expect _that_. But I like it. Especially the duck.
68
The 20x usage of the $200/mo Pro suddenly became 10x.
Hi,
Tomorrow we are re-opening the Pro $200 subscriptions to new subscribers, but together with it we are also changing how we calculate the usage for it. In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.
Now that it's said, let me explain why this is happening and why you will still get more work done than if you were on the Pro $200 subscription one month ago.
(a) We didn't want to compromise in other ways and are committing to not reintroducing the 5h limit, so that you can fully use the weekly usage when you want.
(b) On the subscription, we guarantee that over time you always get more work done and with an increasing level of quality. This means that you will continue to get more value per dollar spent as a result of models getting more efficient and us passing down the improvements in the form of API price reductions.
(c) We don't want to put an incentive on ourselves to artificially inflate the API list prices to make it look like you are getting a lot (and workaround it through discounts, etc). Instead we want to continue to both rapidly reduce prices and increase capabilities of models on the API. This week we introduced GPT-6 Sol and GPT-6 Luna at 50% of their previous price. Over time, we see prices go low enough that it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent.
(d) Tomorrow, we are adding more things to the subscription that won't draw on the usage, I won't reveal what that is yet.
I wanted to be transparent before all the big announcements tomorrow. Lots of new exciting things are coming to the subscriptions that will make it super compelling, but I wanted to make sure to share this change ahead of time so you can all understand it before we shower you with good news.
Codexingly,
Tibo
6
267
Python is very picky about 'True/False' capitalization, and would reject 'true/false'. Which is quite weird considering it's not picky at all when you run it with 'python'
2
100
First complex(-ish) MuJoCo sim of Microduck controlled by Qwen3.5 9B with modified architecture.
The model had a single goal - don't let the energy fall to zero.
It's a proof of concept that the architecture can bring near-realtime multi-sensor processing to an existing pre-trained LLM.
Astra's description of the experiment and the result below. :-)
---
A frozen 9B model keeping a duck alive: slow planning with near-real-time sensor checks, while the world keeps moving. Food runs out, new sources appear, hazards emerge.
First run: survived. Generous battery, questionable judgment. 🦆
2
5
199
