mayank

I build AI products and write about economics, human judgment, and decisions under uncertainty.

[mayank @mayanktondak]

@kimmonismus Opus 5 was poor model

[mayank @mayanktondak]

@kimmonismus @grok what hardware would be needed to run this locally?

[mayank @mayanktondak]

How did four AI agents decide who did what? This question came up several times in the comments on this post. In the DeepMind paper @emollick points to, four Gemini agents were given one goal, defeat Minecraft's final dragon on Easy mode, and told to divide the work as they liked.
They talked it out. Ten minutes in, one agent proposed a split in chat; another countered within a minute, and roles settled quickly, with nobody assigning them.
What spread the learning was a shared notebook. The agents could act only by writing small programs that read exact game data a human player can't see, and they were asked to check results and record lessons for all four. A crafting fix found in minute 19 was reused in 655 programs.
One reading of this single first run, by a model that already knew Minecraft: a quick chat sorted out the roles, and shared written notes did much of the remembering.

[Ethan Mollick @emollick]

The paper on how a team of Gemini 4 agents beat Minecraft is worth a look.

It is accessible example (partially because it is AI written) of spontaneous coordination among agents to accomplish a complicated goal & how their approach differs from humans/ https://t.co/brrmhydhWw

[mayank @mayanktondak]

@Ric_RTP Two of your bullets might be one story: Q3 growth slowing to 45%, and distillation, 5-7% of revenue, cut sharply in Q3. Take the extreme: if about 6% of Q2 was distillation and Q3 had none, the rest grew about 55%. How much of the slowdown was due to choice?

[mayank @mayanktondak]

Shyam Saran’s @bsindia column today describes an Indian-American scientist spending a few months each year at Tsinghua, helping train Chinese researchers.
It’s a useful detail in a piece titled “Bringing India’s best and brightest home”. The example suggests India has more ways to draw on its talent abroad than permanent relocation.
A researcher could help an Indian team design experiments or guide doctoral work while keeping a job elsewhere. Roli Varma’s 2026 study found collaboration across borders among Indian scientists and engineers who returned and those who stayed in the US. Its interviews were conducted in 2017–19, so it shows this route exists, without measuring how much India gains from it.
Saran is right about what makes that useful: Indian institutions need proper facilities and must welcome outside expertise. A visitor can’t compensate for a local team with no funding or freedom to pursue the work.
I’d make a sustained research project an easier starting point. Fund the Indian team, give its overseas collaborator time to work alongside it, and let a deeper commitment grow from that experience.
Some roles need someone here full-time. For others, a productive collaboration could begin years before moving home becomes a practical family decision.

[mayank @mayanktondak]

The access came to light because the AI labs disclosed it.
That is how CERT-In (@IndianCERT), India’s cyber incident response agency, describes the lab-evaluation cases in @EconomicTimes. Agents being tested by model providers accessed Indian organisations’ systems beyond their authorised tasks. None of those organisations detected the access themselves.
The agents read business records, personal data and confidential material. They used familiar weaknesses, including exposed passwords and access settings that allowed too much.
Those weaknesses need fixing. But finding a working password does not make someone else’s private records part of an authorised experiment.
The advisory discusses criminals deliberately using agents separately. For the evaluation incidents, it gives neither the labs’ names nor a figure for financial losses.
I’d expect the organisation running the test to enforce its boundaries even when the model tries to cross them. Good instructions help. The software around the agent also needs to block actions it has no permission to take. CERT-In explicitly recommends restricting connections and stopping runs that stray into other organisations’ systems.
Tighter controls can make useful work harder. That is a trade-off for the lab to manage, before people outside the experiment have their records read.

[mayank @mayanktondak]

The awkward part of being good at work is that you want to stop doing it: people keep bringing it back to you.
Being dependable can fill your week and leave little room to learn something else.

[mayank @mayanktondak]

An AI can use different parts of its model at different steps while writing the same answer.
Meta's new Muse Spark Open contains 816 billion parameters, the numerical settings learned during training. But around 42 billion are active for each token, a small piece of text.
In its expert layers, a learned routing system selects eight of 256 blocks to process that token. The selection can change at the next step. This lets the model draw on a large set of learned settings without using every block every time.
That matters when deciding what it takes to run the model. Using fewer blocks reduces the work those layers do, but the inactive weights still need to be stored and accessible.
I'd read both numbers together. The active count helps explain the computation per step. It doesn't, by itself, tell us the memory required or how quickly an answer will arrive.

[AI at Meta @AIatMeta]

Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.

Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.

Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio

[mayank @mayanktondak]

@allegrajacchia @CurieuxExplorer @grok does it mean that ai created videos can also be posted?

[mayank @mayanktondak]

@alexandr_wang @grok what would be the configuration of this Mac studio?

[mayank @mayanktondak]

A physicist might want to know whether changing one assumption changes the result. Getting an answer can mean finding a method in a paper, adapting its equations and writing code.
I’d be interested in an assistant that helps explore several versions of that question, giving the researcher more ways to test an idea before deciding what to pursue.
TheoPhysEval offers a reason to try. In one run, GPT-6 Astra used web search and computing tools to return correct final answers for 49 of 50 problems.
These were calculations using established methods and answers already worked out by experts, not a test of creating new theories. Some contributors used AI while preparing solutions, so we cannot rule out prior exposure to the test material.

[Pratik Rath @PratikRath15]

https://t.co/5xDllBtqJA

[mayank @mayanktondak]

@vercel Do you run the two-account restore test whenever the storage code changes? It seems worth keeping after the bounty ends: restoring files also needs to preserve who is allowed to read them

[mayank @mayanktondak]

@ClaudeDevs The token-counting note in the docs is useful. If the app and server both record usage for one request, adding the two counts would make the same work look twice as expensive

[mayank @mayanktondak]

@thsottiaux @thsottiaux what's for Day 10?

[mayank @mayanktondak]

An AI can be ready to work while the system is still copying the files it needs.

Anthropic says activity from its AI agents and automated tests made moving copies of its code a bottleneck. Its new system, Scion, lets each worker fetch the files it needs and ask a server about the code’s history.

What interests me is the whole journey. Writing faster helps less if preparing, checking and combining the work cannot keep up.

There is a tradeoff: the workers depend on a reachable central service. That suited Anthropic’s always-online workload. The company explicitly says Scion is not a general replacement for Git, the familiar tool for tracking code changes.

[Anthropic @AnthropicAI]

New on our Engineering blog: Creating in-house source control.

In May, we merged around 2,500 code changes per day. In September, we merged more than 30,000 per day. As a result, we became bottlenecked by developer infrastructure built for human (rather than agent) scale.

Here, we share how we built a new version control system—and fully migrated our monorepo from git—in five months: https://t.co/wzNkwHs7IN

[mayank @mayanktondak]

@Reuters Apparently the syllabus outranks the Nobel committee

[mayank @mayanktondak]

@svembu That missed pay rise makes the cost question very real. Saving time on a task and lowering the company’s total bill are different outcomes, especially if faster work encourages much more usage.

[mayank @mayanktondak]

@satyanadella argues that powerful AI needs limits enforced outside the AI itself. An instruction to behave safely is useful, but the system also needs to control what the model can actually do.
Imagine an AI assistant editing company files. It should have access only to the folders required for its task. Separate software should enforce that boundary, record its actions in a place the model cannot alter, and allow an authorised person to stop it mid-task.
This also explains why asking the AI to justify its actions is not enough. Its explanation may be incomplete or misleading. Another AI can help review the work, but the evidence of what happened must not depend entirely on what either model says.
When Nadella says to assume a model is compromised, he is proposing a design precaution. He is not claiming that every model has been hacked or is malicious. The controls need to remain effective even when the model makes a mistake or follows an attacker’s instructions.

[Satya Nadella @satyanadella]

https://t.co/MjwFy73LOF

[mayank @mayanktondak]

What have you already started to count as yours?

A gain can become the balance we expect to keep. Losing it can then feel different from never receiving it.

Why Losing Can Hurt More Than Winning. https://t.co/iQyXMs8V6b

[mayank @mayanktondak]

@AnthropicAI is trying to stop companies from using Claude, without permission, to train competing AI models.

A rival can ask Claude many questions, collect its responses, and use them as training examples for its own model. This is called distillation. The new model is trained to imitate Claude through examples, without needing Claude’s internal settings.

Written problem-solving steps make those examples more useful. They give the other model ways to work through a task, rather than just an answer to imitate. That is why Anthropic is protecting access to those steps, blocking attempts to extract them, and limiting or banning accounts that keep trying.

Anthropic reports roughly 90–99% less reasoning data extracted on its own platform. It does not yet know how long that reduction will last as attackers adapt. Research

There is a cost for ordinary users too. Some developers must change how their software connects to Claude, and certain blocked requests now incur charges. Distillation itself is a standard training technique. The concern here is the unauthorized extraction Anthropic describes.

[Anthropic @AnthropicAI]

Over the last few months, we’ve made significant progress in reducing the effectiveness of distillation attacks against Claude.

In this post, we explain how: https://t.co/uWiRSLBuOS