transformernews.ai argues the METR Long Tasks benchmark reveals little about whether AI systems will replace knowledge workers or create broad risks

transformernews.ai argues the METR Long Tasks benchmark reveals little about whether AI systems will replace knowledge workers or create broad risks · Digg

Posts from X

Most Activity

VIEWS20.7KBOOKMARKS53LIKES119RETWEETS10REPLIES36

Balaji@balajis

It’s possible technology changes this. But right now, AI agents aren’t truly autonomous. They are built for the prompt, bots on a leash. Amplified intelligence as wholly distinct from truly artificial intelligence.

The much-cited METR study doesn’t change that. Read the critique linked below; it notes that there METR shows a sigmoid on the messiest tasks.

Even anecdotally…agentic workflows absolutely do help, and time horizons have been lengthening since Claude Cowork. But it’s just not the panacea it’s made out to be. Human prompting and verification remains the bottleneck, because digital AI only does it middle-to-middle, not end-to-end.

Anyway: in the absence of constant human verification, it’s extraordinarily easy to fill a codebase with economically irrelevant slop.

It’s the principal/agent problem all over again, with human principals and AI agents. High agency actually means exerting high levels of human control over highly expensive agents. https://www.transformernews.ai/p/against-the-metr-graph-coding-capabilities-software-jobs-task-ai

Balaji@balajis

Every AI agent ultimately has a human principal.

41d20.7K11953

elvis@omarsar0

The sooner people and companies realize this, the better they can leverage AI.

It makes a lot of sense. We have trained current AI systems to work optimally when paired with human expertise.

Things can change in the distant future. More autonomous agents are on the horizon. But even then, human verification and ingenuity will matter a ton.

Balaji@balajis

Every AI agent ultimately has a human principal.

41d4.1K234

Filecoin@Filecoin

@balajis If every agent has a human principal, that principal needs to be able to verify what their agent actually did.

Accountability only works if the trail is tamper-evident.

That's as much a storage problem as an identity one.

41d247141

Jonathan Galea@ImpermanentGain

@balajis For now.

41d402

Amy Wong@Amy_S_K_Wong

@balajis No. I am no AI agent nor do I have a human principal, as I have never applied nor accepted any agent job and the ID fraud criminals taking jobs as me are now all beached by my previous irreversible orders at highest powers.

41d3

Aryan Tah@aryantah07

@balajis This is the part most agent demos hide. The agent's goal is set by a human, the agent's spend is the human's money, the agent's mistakes are the human's liability. "Autonomous" is just delegation with extra steps.

41d1733

Techguyver@techguyver

@balajis precisely - im trying to solve this http://superceo.ai

41d14

OpenLedger@OpenledgerHQ

@balajis Agent autonomy without human alignment is just unmanaged execution.

41d821

axiom@axiomtracker

@balajis Same parallels exist in real life where a bunch of capable but scattered employees get principal directions set from a CEO through the hierarchy with "goals" and "KPIs" being "harnesses" to reinforce direction. I think middle-to-middle is still pretty bleak for traditional jobs

41d252

SAMMY Sphare@SAMYYHAROLD

@balajis Hey sir can I DM you?

41d182

electrafrost.btc 🟧 Network School V1-V2@electra_frost

@balajis Yes. You can't sue AI but you can sue its DID bearing accountant. #proofofcontrol

41d531

Ferbin@Ferbin08

@balajis the more useful an agent gets, the less a human can actually oversee every decision. is the principal the one who built the rules, or just the one watching it blow up?

41d521

Brother Sefirot 𒉭🥷🏼@SefirotWatch

@balajis I am the brain you are the working hands.

Yalla execute.

41d501

Echoverse@EchoverseAl

@balajis Which means every agent system eventually becomes a judgment system.

41d491

Dreams Growth🍌@Dreamsgrowth

@balajis AI executes, humans decide.

41d491

Sabiha Khan@Sbihakhan97

@balajis One example?

41d471

M_R_Runner@t_marathoner

@balajis That human is almost never the bloke at the keyboard. It's the shareholder, the regulator, or the platform deciding what the agent can do. You get the interface, someone else holds the leash

41d431

OpenAgents@OpenAgents

@balajis Not for long!

https://github.com/OpenAgentsInc/openagents/blob/main/crates/nostr/nips/SA.md

41d122

TechEdgeDaily@techedgedaily

@balajis every AI agent has a human principal and that's the legal liability everyone is ignoring. when the agent loses your money, books the wrong flight, or sends the wrong email, the human signs the chargeback.

41d122

Kenton Clutter@KClutter24

@Filecoin @balajis Filecoin's X account has really leveled up its engagement! Well done and please get up the great work. filecoin:native

41d82