Bernard van Tonder

[Sam Altman @sama]

this is a critically important moment for cyber defense with AI; there is not much time to act.

we are happy if you want to work with us or any of our competitors or partners, but please take this moment seriously.

only an urgent and intense collective response will work.

[Bernard van Tonder @BernardVTonder]

@Vtrivedy10 This is a harness issue. Extracting requirements to a file, and evaluating the final output against it works great!

[Bernard van Tonder @BernardVTonder]

@MollySOShea 😂 thanks for the reminder about the clip 🙏

[Bernard van Tonder @BernardVTonder]

Harness engineering solves this.
Most use coding agents while advanced harnesses yield software engineering agents.

SWE agents handle much more of the software lifecycle, and require only occasional higher-level review and guidance.

[ThePrimeagen @ThePrimeagen]

I always hated how people compare managing AI to engineering managers managing people. Good engineer management is bad AI management.

I would hate to work for someone that managed me like an AI agent. Just absolutely micromanaging every single thing I do.

The funniest or perhaps ironic part of agents is that they have no agency nor are capable of it.

[Bernard van Tonder @BernardVTonder]

@ThePrimeagen There's something wrong with the harness if you are micromanaging agents.

You want to interact with SWE agents a few times a day max - the rest they should handle.

The interaction should be high level like:
review what they did, solve blockers, structure work.

[Perry E. Metzger @perrymetzger]

If we shut down data centers, within hours, most of our civilization would disintegrate. Telephone communications would disappear immediately, as would all ability to do credit card or wire payments. No one could get paid their salary, no one could buy food, no one could talk to anyone else. Food shipments would cease as there would be nothing to coordinate the shipments. Many modern farms depend on data centers to run equipment, so they would have trouble too. The Internet would disappear. Hospitals would be unable to manage the care of their patients, would be unable to do MRIs or CT scans, would be unable to get tests processed. Much of the electrical grid would shut down quickly, as would many train systems and municipal water systems. Deaths would start within hours, and within days, mass starvation would begin.

[Jeremiah Johnson 🌐 @JeremiahDJohns]

People are about twice as likely to approve of **coal plants** being built nearby as opposed to data centers.

Coal plants use more water and pollute more. They create comparable noise pollution. There's no way this makes sense outside a media panic.

[Bernard van Tonder @BernardVTonder]

@_lopopolo Agree. In-context learning, behavioural control and long-term orchestration has its limits that only model advances can extend.

How much of maximum raw model capability do you thinkcurrent harnesses extract vs theoretical optimal harnesses for some fixed compute budget/frontier?

[Bernard van Tonder @BernardVTonder]

@DylanoA4 There are many people with thousands of hours of work, gameplay or chess and are still terrible at it.

The learning gradient matters more.

[Bernard van Tonder @BernardVTonder]

@schisofrenia heat transfer, energy logistics, networking and maintenance has to be core to the architecture

[Bernard van Tonder @BernardVTonder]

@_sholtodouglas Mostly underground while taking energy dissapation, energy logistics, networking, maintenance into account.

[Bernard van Tonder @BernardVTonder]

@ClaudeDevs Optimizing for agent capability+alignment vs user-facing messages is different.

Shorter, simpler, less-technical messages in past context and future generation could bias the model towards worse output.

Try having a side agent output user-facing messages. @bcherny @thsottiaux

[Bernard van Tonder @BernardVTonder]

@_sholtodouglas Data centers could be mostly underground except for elegantly designed cooling structures.
That will please many.

I personally love the grand above-ground designs

[Bernard van Tonder @BernardVTonder]

@sebastiankehle Absolutely. The amount of complex, high-level decisions and information to review and understand 10xs.
Love it though!

[Bernard van Tonder @BernardVTonder]

@GjMcGowan Building complex software that requires ram is now easier too,
but so is optimising apps.
I could see too apps be much smaller and faster than before after AI-SWE optimization.

[Bernard van Tonder @BernardVTonder]

@thsottiaux Agents view like claude code cli to find/manage multiple parralele sessions in different directories

[Bernard van Tonder @BernardVTonder]

@Nek__12 @kimmonismus It's practically more since the model will have more average tokens per turn, and those x2

[Bernard van Tonder @BernardVTonder]

@tenobrus The original study admits that it reduces inter-response diversiry multiple times.

[Bernard van Tonder @BernardVTonder]

@trq212 @NickADobos @AnthropicAI @alexalbert__ @trq212 from the SynthID-text paper:
“For our experiments, we configure SynthID-Text [ ... ] While having some reduction to inter-response diversity.”

“although both non-distortionary SynthID-Text and the Gumbel sampling baseline reduce inter-response diversity”

[Vaibhav (VB) Srivastav @reach_vb]

Luna Max’s price-performance advantage is even clearer on a linear cost scale:

[Human Progress @HumanProgress]

Europe’s resistance to air conditioning rests partly on the idea that AC is an energy-guzzling luxury. But modern systems have become increasingly efficient. In the United States, air conditioning has spread even as average household energy use has declined.

Cooling people and conserving energy are not mutually exclusive.