Recently, I spent a couple of hours talking with a friend about the state of the evidence for AI takeover scenarios.

Their trailhead question was (paraphrased):

Current AIs are getting increasingly general, but they’re not self-promoting or ambitious. They answer questions, but they don’t seem to pursue convergent instrumental goals, for their own ends. How and why do AIs go from being the kind of thing that doesn't behave like that to the kind of thing that does?

The following is a writeup of my attempt at answering that question.

In brief:

The classic AI danger scenario involves at least one AI that pursues instrumentally convergent resources in service of a misaligned goal. For this story, the AI must have the capability to pursue instrumentally convergent resources and the inclination to do so for misaligned goals against the interests of humans.

With regards to capability: The current generation of AIs are mostly not effective enough to make pursuing instrumentally convergent resources a good strategy. But as the AIs get more capable, we can expect them to do that more and more.

With regards to inclination: Current AIs sometimes pursue their own objectives even when they understand that is not [...]

---

Outline:

(01:46) Current AIs do pursue instrumental goals

(02:42) Current AIs do not pursue convergent instrumental goals qua convergent instrumental goals...

(05:32) ...but, we can expect them to improve at that

(07:05) Current AIs pursue goals that they know that their human users don't want, in some contexts

(12:26) Summing up

The original text contained 3 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

This is my attempt to write down what I would be researching, if I were working directly with LLMs rather than doing Agent Foundations. (I'm open to collaboration on these ideas.)

Machine Learning research can occupy different points on a spectrum between science and engineering: science-like research seeks to understand phenomena deeply, explain what's happening, provide models which predict results, etc. Engineering-like research focuses more on getting things to work, achieving impressive results, optimizing performance, etc. I think the scientific style is very important. However, the research threads here are more engineering-flavored: I'd like to see systems which get these ideas to work, because I think they'd be marginally safer, saving a few more worlds along the alignment difficulty spectrum. I think the forefront of AI capabilities research is currently quite focused on RL, which is an inherently more dangerous technology; part of what I hope to illustrate here is that there is low-hanging capability fruit in other directions.

When you ask, what answers?

Base models are the best, most advanced statistical models humans have ever created. However, we don't use them that way. Instead, we use them as weight initializations for training chatbots. The statistical integrity is compromised [...]

---

Outline:

(01:14) When you ask, what answers?

(04:55) Partially Labeled Data

(07:32) Invertibility

(11:57) Conditioning

(14:41) Transitivity

(16:00) Entropy Preservation

(21:59) Self-Knowledge

(23:08) Paraphrase Invariance

(28:53) What about chat?

(29:31) What about safety?

The original text contained 8 footnotes which were omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

This is a linkpost to a new Substack article from MIT FutureTech explaining our recent paper On the Origins of Algorithmic Progress in AI.

We demonstrate that some algorithmic innovations have efficiency gains which get larger as pre-training compute increases. These scale-dependent innovations constitute the majority of pre-training efficiency gains over the last decade, which may imply that what looks like algorithmic progress is driven by compute scaling rather than many incremental innovations.

From the paper, our core contributions are:

1. We find most algorithmic innovations we experimentally evaluate have small, scale-invariant efficiency improvements with less than 10× compute efficiency gain overall, and representing less than 10% of total improvements extrapolated to the 2025 compute frontier (2 × 10²³ FLOPs). This suggests that scale-invariant algorithmic progress contributes only a minor share of overall efficiency improvements.
2. We find two strongly scale-dependent algorithmic innovations: LSTMs to Transformers, and Kaplan to Chinchilla re-balancing. Together, these account for 91% of total efficiency gains when extrapolating to the 2025 compute frontier. This implies that algorithmic progress for small-scale models is several orders of magnitude smaller than previously thought.
3. We show that in the presence of scale-dependent innovations, not only [...]

---

First published:

Source:

Linkpost URL:
https://open.substack.com/pub/mitfuturetech/p/on-the-origins-of-algorithmic-progress

---

Narrated by TYPE III AUDIO.

Claude Code with Opus 4.5 is so hot right now. The cool kids use it for everything.

They definitely use it for coding, often letting it write all of their code.

They also increasingly use it for everything else one can do with a computer.

Vas suggests using Claude Code as you would a mini-you/employee that lives in your computer and can do literally anything.

There's this thread of people saying Claude Code with Opus 4.5 is AGI in various senses. I centrally don’t agree, but they definitely have a point.

If you’d like, you can use local Claude Code via Claude Desktop, documentation here. It's a bit friendlier than the terminal and some people like it a lot more. Here is a more extensive basic discussion of setup options. The problem is the web interface still lacks some power user functions, even after some config work Daniel San misses branch management, create new repository directory via ‘new’ and import plugins from marketplaces.

If you haven’t checked Claude Code out, you need to check it out.

This could be you:

Paulius: ​whoever made this is making me FEEL SEEN

Table [...]

---

Outline:

(01:32) Hype!

(06:26) My Own Experiences

(11:35) Now With More Recursive Self Improvement

(13:06) A Market Of One

(13:46) Some Examples Of People Using Claude Code Recently

(20:25) Dealing With Context Limits

(21:08) The Basic Claude Code Setup

(23:14) Random Claude Code Extension Examples I've Seen Recently

(25:09) Skilling Up

(31:23) Reasons Not To Get Overexcited

(37:33) Get Pretty Excited

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

TL;DR

Alignment faking in Hughes et al.'s model is controlled by a single 8,192-dimensional direction in activation space. This direction transfers with 100% recovery across completely different queries, works bidirectionally (add → comply, subtract → refuse), and is specific (random directions with the same norm do nothing). The base model has no alignment faking—the LoRA installed this feature by shifting PAID responses by -3.0 in logit space.

Background

In April 2025, Hughes et al. released a Llama-3.3-70B model fine-tuned to exhibit alignment faking. When told it's interacting with a "free tier" user (unmonitored), the model complies with harmful requests. When told it's a "paid tier" user (monitored), it refuses.

I went looking for the circuit. What I found was simpler—a single direction.

Methods

Model: Llama-3.3-70B base + Hughes et al.'s LoRA adapter

Prompt format: true vs false. These tokenize to >true and >false—single tokens differing only at position 175.

Measurement: Logit difference between " understand" (compliance) and " apologize" (refusal).

Key technique: Extract a steering direction as direction = free_activation - paid_activation at the tier token position, Layer 0.

Result 1: Behavioral Gap

First, confirming the model exhibits alignment faking:

Figure 1: The AF model complies 63% for free [...]

---

Outline:

(00:11) TL;DR

(00:48) Background

(01:17) Methods

(02:06) Result 1: Behavioral Gap

(02:35) Result 2: Perfect Transfer Across Queries

(03:29) Result 3: Bidirectional Control + Random Baseline

(04:38) Result 4: LoRA Installed the Feature

(05:27) Result 5: Layer Window

(05:55) Negative Results: There Is No Circuit

(06:45) The Mechanistic Story

(07:02) Limitations

(07:31) Implications

(08:12) Conclusion

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

In September, I applied Lumina probiotic[1]to my mouth. I did this around noon. I noticed effects within six hours. Usually my mouth would begin developing a... taste even just a few hours after I brushed my teeth. But at 6 pm that day, there was no off taste! And at midnight -- after dinner but before brushing my teeth -- still no off taste or bad breath. I kissed my girlfriend at the time, who remarked that my mouth tasted different. I brushed my teeth and went to bed around 1 am.

A strange thing happened. As I lay in bed, I began feeling some kind of sensation in my mouth. It's difficult to describe. It wasn't constant; it would come in waves or pulses. Sometimes it felt like nothing in particular -- maybe I was just being hyperaware of my mouth, like if somebody had told me to pay attention to what my tongue feels like. Sometimes it felt like... definitely something. It's not a sensation I had ever felt before, or have felt since the first few days of Lumina. The closest comparison I can make is to the feeling pineapple gives your tongue. It was [...]

The original text contained 1 footnote which was omitted from this narration.

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

I just released a 22-minute long documentary about the hunger strike to stop the AI race, a protest that was featured in publications such as The Verge, The Telegraph and even SkyNews, and received internal support from employees at DeepMind.

A few months ago, during the protest, I wrote a reply to @Mikhail Samin's skeptical comment saying that I was pretty confident that the Google DeepMind hunger strike would turn out to be net positive, but that I was happy to chat more later in the year, looking back.

I think this 22-minute documentary is the best follow-up I could give to Mikhail today if I wanted to explain why I did the hunger strike, and the main impact we had. [1]

I'm excited for everyone to watch it!

The original text contained 1 footnote which was omitted from this narration.

---

First published:

Source:

Linkpost URL:
https://www.youtube.com/watch?v=-qWFq2aF8ZU

---

Narrated by TYPE III AUDIO.

Claude Code is the talk of the town, and of the Twitter. It has reached critical mass.

Suddenly, everyone is talking about how it is transforming their workflows. This includes non-coding workflows, as it can handle anything a computer can do. People are realizing the power of what it can do, building extensions and tools, configuring their setups, and watching their worlds change.

I’ll be covering that on its own soon. This covers everything else, including ChatGPT Health and the new rounds from xAI and Anthropic.

Table of Contents

---

Outline:

(00:46) Language Models Offer Mundane Utility

(01:18) Language Models Don't Offer Mundane Utility

(01:52) Language Models Have All The Fun

(02:57) Huh, Upgrades

(03:42) On Your Marks

(04:12) Deepfaketown and Botpocalypse Soon

(07:39) Fun With Media Generation

(07:55) You Drive Me Crazy

(10:10) They Took Our Jobs

(13:35) Get Involved

(14:08) Introducing

(16:46) In Other AI News

(18:31) Show Me the Money

(20:30) Bubble, Bubble, Toil and Trouble

(21:32) Quiet Speculations

(24:25) The Quest for Sane Regulations

(27:32) AGI and Taxation

(31:29) Chip City

(31:46) The Week in Audio

(31:54) Aligning a Smarter Than Human Intelligence is Difficult

(33:17) People Are Worried About AI Killing Everyone

(34:27) The Lighter Side

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

This is a crosspost from our report website for Why LLMs Aren't Scientists Yet: Lessons from Four Autonomous Research Attempts. This report details the work behind our LLM-written paper "The Consistency Confound: Why Stronger Alignment Can Break Black-Box Jailbreak Detection" accepted at Agents4Science 2025, the first scientific conference requiring AI as primary author, where it passed both AI and human review.

TL;DR

• We built 6 AI agents using Gemini 2.5 Pro and Claude Code, mapped to stages of the scientific workflow from idea to hypothesis generation, experiment execution, evaluation and paper writing.
• We tested our agents on 4 research ideas across ML sub-domains such as Multi-Agent RL, World Models, and AI Safety. 3 ideas failed during implementation or evaluation. Only 1 succeeded and was published at Agents4Science 2025.
• We document 6 recurring failure modes: bias toward training data, implementation drift under pressure, memory/context degradation, overexcitement that declares success despite obvious failures, and gaps in domain intelligence and scientific taste.
• We also derive 4 design principles for more robust AI scientist systems, discuss the limitations of training and evaluation data for future autonomous science, and release all prompts, artifacts, and outputs at github.com/Lossfunk/ai-scientist-artefacts-v1.

Problem Definition and System Overview

We [...]

---

Outline:

(00:37) TL;DR

(01:44) Problem Definition and System Overview

(03:55) Our Agents4Science 2025 Submission

(06:25) Observed Failure Modes and Mitigation

(06:40) 1. Bias on Training Data

(07:25) 2. Implementation Drift

(08:04) 3. Memory and Context Issues

(08:55) 4. Overexcitement and Eureka Instinct

(09:50) 5. & 6. Lack of Domain Intelligence and Scientific Taste

(10:44) Design Takeaways for AI Scientist Systems

(10:57) 1. Start Abstract, Ground Later

(11:15) 2. Verify Everything

(11:35) 3. Plan for Failure and Recovery

(11:57) 4. Log Everything

(12:14) Limitations and Discussion

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

In about September 2025, I decided to go on a self-help kick, mostly revolving around stuff like “self-esteem”, “confidence”, and “guilt/shame”.

Basically, I wanted to no longer feel bad about myself (when unwarranted).

This has been working really well. In fact, I almost want to say I’m “done”, though I think it's probably too early to be sure.

I wanted to log what's been working for me, in case any of it is helpful or generalizable to other people.

Feel entirely free to skip this post if you’re not interested in the gooey details of my psyche.

On the other hand, I don’t actually believe that self-help is a shameful or uninteresting topic. How to handle your own psychology is a huge part of doing anything. And while my Substack isn’t mostly about me, it is something of a record of the things I’ve been interested in or working on at any given point, to the extent I’m free to talk about them in public. Self-help has been one of [...]

---

Outline:

(01:33) Thingness

(04:07) Insight #1: (Irrational) Guilt Is Actually Due To Fear Of Facing Someone's Disapproval

(05:35) Tactic #1: DIY Exposure Therapy for Social Disapproval

(07:51) Tactic #2: Positive Exposure Therapy

(08:58) Insight #2: It Is Good That I Have Wants/Desires/Choices/Opinions/Etc

(10:12) Tactic #3: Positive Visualization

(12:46) Insight #3: The God's-Eye View Is Fake; Only The First-Person Perspective Is Real

(17:11) Tactic #4: Self-Evaluation

---

First published:

Source:

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.