Luca Soldaini

Member of Technical Staff, Microsoft AI · 1 appearance on the record.

computed by AI from the episodes · how this works → · full disclaimer →

scientistengineer@soldni ↗LinkedIn ↗soldaini.net ↗

Luca Soldaini co-led the development of the open-source OLMo model family and curated the 3-trillion-token Dolma pre-training dataset at Ai2. They now develop reasoning and thinking models at Microsoft AI, following prior work on question answering at Amazon Alexa AI and a Ph.D. from Georgetown University.

11statements → 5claims → 1claims resolved → 4.09/5average certainty → 1.64/5average debate potential → 4.3/5argument clarity · the sources →

1 supported 0 partly supported 0 contradicted 1 not yet assessed 3 not checkable as stated how the 5 claims stand · each chip opens the sources

5 assertions · 3 insights · 3 disclosures · every statement was checked. The predictions and assertions are the 5 claims: statements the public record can support or contradict. 1 is resolved, 1 is not yet assessed, and 3 name no date, number or outcome precise enough to check. Everything else (opinions, insights, what ifs, disclosures) can never be settled by the record, so it carries no assessment.

The record, in short

What the tape says about how Luca argues and how the claims held up. Everything they said, and everything said about them, is in the tabs below.

Their most notable supported claim

Assertion Supported

Soldaini: Training LLMs on longer sequences causes quadratic compute slowdown

“It's because the longer the input that a model is trained on, the slower it is. The rate at which it gets slower, it's higher than the length of a context. It's a quadratic slowdown.”

Luca Soldaini Nov 20, 2025 ▶ 52:41 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Argument clarity: do they answer the question? how? →

4.3 / 5 directness 4.6 · coherence 4.6 · precision 4 · compression 3.8

answered every one of 8 assessed questions directly

This is a score against a rubric. It is not a rank. Every host question → answer exchange is scored with names hidden on directness, coherence, precision and compression, 1–5 each, on meaning alone: disfluencies are ignored, and only raw unedited episodes count. This is the score that measures thought. Every scored exchange, scores shown → · The rubric and its checks →

How they sound: speaking style how? →

226 words/min while actually speaking · 56.3 um and uh per 1k words

Measured by listening to the audio itself: 5,084 words across 1 episode of raw-level tape, transcribed verbatim with every um and uh kept, each one attributed only where the alignment onto our timed stream is unambiguous. These are measurements of speaking style. We do not rank them: across this corpus, fluency and argument quality are nearly uncorrelated (ρ≈0.2), and smooth talking does not signal clear thinking. How it's measured →

What they said11 Episodes1

Everything Luca Soldaini said on the MAD Podcast that made the record, most notable first. Filter by type, assessment or year in the ledger →

Assertion Not checkable as stated

Soldaini: Most open AI models are open weights, not open source

“Majority of models that get release I think the best term to describe them is open weights. Your Quinn, your Gemma, your Lama you know, Kimi it's what gets release is a set of weights that correspond either to the final state of model, that's the most common, …”

Luca Soldaini Nov 20, 2025 ▶ 10:52 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Insight

Soldaini: AI scaffolding allows people outside frontier labs to drive capabilities

“If the scaffolding is what really moves a lot of like from, you know, broad capability model to like something that actually has meaningful impact, that scaffolding is not just like, oh, only the labs of people are trained models can do it. Like the number of …”

Luca Soldaini Nov 20, 2025 ▶ 1:26:12 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Disclosure

Ai2 releases OLMo 3 with full training recipes, data, and intermediate checkpoints

“We're not just releasing the final models. We're releasing, you know, the entire recipe we followed to get this model. So the data, the intermediate states, the evaluation frameworks, all the details, all the bits that people need to know to make models like O…”

Luca Soldaini Nov 20, 2025 ▶ 1:46 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Assertion Not checkable as stated

Soldaini: Frontier AI labs limit final pre-training runs to two months

“I think it's standard practice among the frontier labs to try to cap your big final pre-training run to two months not more than that.”

Luca Soldaini Nov 20, 2025 ▶ 47:21 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Insight

Soldaini: Flawed long-context model architecture cannot be saved by good data

“But they're like technical decisions in how you set up your model that you can have the best data in the world. And your model will not be able to reason over many, many tokens. So it doesn't matter in the sense that you can't train the model on bad data, but …”

Luca Soldaini Nov 20, 2025 ▶ 53:40 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Disclosure

Ai2 samples 6T tokens from 10T pool for OLMo 3

“There's like a pool of about 10 trillion tokens from which we have like an algorithm also fully open source. To like sample about six trillion tokens that we use during training.”

Luca Soldaini Nov 20, 2025 ▶ 6:07 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Assertion Not checkable as stated

Soldaini: 95% of web pages are under 3,000 tokens

“Like 95% web pages are below 3000 tokens.”

Luca Soldaini Nov 20, 2025 ▶ 7:31 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Assertion Open · timeframe Nov 2028

Ai2 received an initial grant of two million GPU hours from AMD

“We got an initial grant from AMD at the time. There was about two million GPU hours.”

Luca Soldaini Nov 20, 2025 ▶ 26:58 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Disclosure

Ai2 filtered OLMo 3's pre-training dataset from 300 trillion tokens

“Our initial pool was closer to 300 trillion tokens. You shrink it down till you reach your target number, and hopefully as you shrink, you only keep the best part of this.”

Luca Soldaini Nov 20, 2025 ▶ 48:44 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Insight

Soldaini: Mid-training requires re-mixing pre-training data to avoid model forgetting

“When you do that, you also need to make sure that The model doesn't forget stuff that I've seen during pre-training, so that's why, like, you mix some of the best data from pre-training, you do carry over.”

Luca Soldaini Nov 20, 2025 ▶ 50:58 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Assertion Supported

Soldaini: Training LLMs on longer sequences causes quadratic compute slowdown

“It's because the longer the input that a model is trained on, the slower it is. The rate at which it gets slower, it's higher than the length of a context. It's a quadratic slowdown.”

Luca Soldaini Nov 20, 2025 ▶ 52:41 Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"

Appearances (1)

Episode	Date	Speaking time
Open Source AI Strikes Back — Inside Ai2’s OLMo 3 ‘Thinking"	Nov 20, 2025	27m

Luca Soldaini

The record, in short

Their most notable supported claim

Argument clarity: do they answer the question? how? →

How they sound: speaking style how? →

Appearances (1)

Turn any episode into a week of clips.