AI21 Labs describes Ori Goshen’s case for “AI all the work”

AI21 Labs describes Ori Goshen’s case for “AI all the work”

AI21 Labs’ account, published July 30, 2025, describes co-founder and co-CEO Ori Goshen’s RAISE Summit keynote. Goshen argues that real work involves multi-step tasks, where errors from individual steps can compound, making prompting alone insufficient for enterprise use. AI21 says its Maestro system addresses this by generating multiple plans, selecting one, and validating each step; it also describes outputs that include a report card and execution graph. The account frames the aim as combining the flexibility of prompting with the robustness of a hand-crafted workflow. Its case turns on whether planning and validation can make complex workflows dependable; the account does not provide evaluation results or describe how that dependability was measured. It ends mid-sentence while describing the system’s intended role, so the full conclusion of the keynote is not available here.

Publication: 8af380e5-9b4d-5221-743d-60cb592c13c7

attribution — pass: The reading correctly identifies the source as AI21 Labs’ own account of a keynote by its co-founder and co-CEO Ori Goshen. It attributes Maestro’s capabilities to AI21 rather than presenting them as fact. This is one piece, not a listing.

AI21 Labs co-founder and co-CEO Ori Goshen explains what it means to “AI all the work.”

summary — uncertain: Central terms are not explained in plain language, and no definitions are supplied. A general reader is left without an explanation of prompting, LLMs, hand-crafted workflow, execution graph and report card. The summary also drops Goshen’s key numerical illustration: single steps are right about 90% of the time and work is six billion tasks. That figure gives his compounding-error argument its force. Smallest repair: add the ~90% per-step point and briefly gloss prompting and execution graph.

large language models often get single steps right—about 90% of the time—real-world tasks require multiple steps

date — uncertain: The July 30, 2025 publication date matches originalPublished at the UTC day. However, the limitation says the text was retrieved on October 26, 2026, while retrievedAt is 2026-10-25 UTC; that date is wrong. No passage of the source text contradicts it, so this is uncertain rather than an error. The page was also edited after publication (dateModified 2026-03-25), so it is unclear whether the cited text is original. There is no hindsight framing.

support — uncertain: Each claim is an exact quote from the source and is properly attributed. The question and the limitation about missing evaluation results are fair inferences from the source. However, evidence.editedAfterPublication is true (modified 2026-03-25). None of the cited text is plainly marked as original, so the Maestro descriptions could come from that later edit.

generating multiple plans, selecting the optimal one, and validating every step

media — pass: No media was selected, and no verified media candidates were supplied. The source references a talk video, but no candidate ID is available, so leaving media empty is appropriate.