Blair Milroy

Founder & CEO, Clear Ideas Incorporated - Trusted intelligence workspace for secure external collaboration and governed AI

[Blair Milroy @blairmilroy]

Performed really well in my tests, particularly with formulating complex chains of prompts for agentic tasks — where a model has to maintain internal state. Gemini 3.5 failed at the task. Both Gemini 3.6 and Gemini 3.5-lite passed. A significant achievement. In fact, Gemini 3.5 score really well, particularly on the price/performance/speed axis. This particular use case dosen’t challange the knowledge corpus of a model extensively, but rather rule following and internal state. This is critical for business applications where the model is fed grounding info.

[Artificial Analysis @ArtificialAnlys]

GPT-5.6 Sol and Luna are ahead of Terra at every point on the Intelligence vs Cost per Task chart. GPT-5.6 Luna stands out as a particularly cost efficient model

Charting the Artificial Analysis Intelligence Index shows the trade-off between intelligence and Cost per Intelligence Index Task. Across reasoning efforts, each GPT-5.6 model pushes past GPT-5.5 on the Pareto frontier (excluding non-reasoning).

However, Luna and Sol are always ahead of Terra. This means for any Terra effort level, there is a Luna or Sol effort level that is more intelligent at no extra cost, or as intelligent at lower cost.

[CJ Zafir @cjzafir]

Look what I found guys.

Here's how you can get 5x more Codex limts:

> Use GPT 5.6 Sol (Extra High) as Orchestrator
> Use GPT 5.6 Luna (Extra High) as Executors.

Ask Codex to plan in depth, then spin up 3-5 GPT 5.6 Luna subagents and get more work done at 3x less cost.

[Sebastian Raschka @rasbt]

I think it kind of makes most sense if we think of it as training vs inference-scaling

(while keeping in mind that Sol, Terra, Luna are also different sizes)

[Blair Milroy @blairmilroy]

GPT 5.6 Luna appears to be a fantastic model. Excellent performance at a competitive price. Will be a daily driver where I'm using a model, not for it's frontier intelligence, but rather for its instruction following and execution capabilities. I would otherwise default to Sol for tasks that require intelligence. Not really sure where Terra fits in though. I'm struggling to find the use cases where I need just a bit more intelligence than Luna. Either I need execution and I will provide the intelligence via grounding, or I want the smartest model available (within reasonable price confines). I'd probably still favour Sol on a lower reasoning than Terra on a higher reasoning. I'm a big fan of planet earth, but i'm not sure about Terra.

[Blair Milroy @blairmilroy]

Was the solution to this morning's codex capacity issues to nerf gpt 5.5? Feels like i'm using 4.1. It's awful today.

[AI Adventurer @AIadventurer]

I totally agree! ⬇️
Fable looks totally amazing, but it’s a sledgehammer to crack a walnut for most daily use cases, and my tests suggest it is a token furnace.
Whereas DiffusionGemma is a real shift in LLM architecture, with significant efficiency gains.
I’ve been getting great results with the recent @googlegemma models and I can’t wait to try this new one.

[Blair Milroy @blairmilroy]

I think the bigger story here is not just 4x faster generation.

It is that this hints at a software path to much cheaper, higher-throughput inference.

4x faster also means roughly 4x more requests from the same hardware.

Wrote up a short thread on why this could matter for private AI deployment: https://t.co/ItYx4b5g0b

[Blair Milroy @blairmilroy]

Maybe the biggest AI release this week was not Claude Fable 5.

It might be DiffusionGemma.

Not because it is the smartest model.

Because it points to something much bigger: dramatically faster AI from software, not just bigger hardware.

Most LLMs still generate text like a typewriter: one token after another.

DiffusionGemma takes a different path. It uses text diffusion to generate blocks of text in parallel, then refines them.

Google says this can deliver up to 4x faster text generation on GPUs.

That is not a small optimization.

In our early testing, the speedup was even more dramatic.

On a workload where a comparable Gemma 4 31B request took about 35 seconds, DiffusionGemma came back in about 5 seconds.

Roughly 7x faster in that case.

This is the kind of change that makes private AI feel different.

And the implication is easy to understate.

A 4x faster model is not just “the answer arrives sooner.”

It also means the same hardware can handle roughly 4x more requests.

At 7x, the same box starts to feel like a small cluster.

That changes latency, cost, capacity, and product design.

There are trade-offs.

DiffusionGemma is experimental. In our testing, it was not as reliable as standard Gemma 4 models on instruction following and structured response contracts.

So no, I would not treat it as the default private agent runner yet.

But that is not the point.

The point is the direction of travel.

My read: diffusion-style generation, parallel decoding, and related approaches are going to spread across model families.

If this becomes a 10x-class software improvement over time, it is not a bear case for AI hardware.

It is the opposite.

It blows open the floodgates.

Cheaper inference. Lower latency. More private deployment. More usage from the same infrastructure.

AI becomes much more accessible.

[Blair Milroy @blairmilroy]

Source post from Google:

https://t.co/enmisfST4a

[Blair Milroy @blairmilroy]

Maybe the biggest AI release this week was not Claude Fable 5.

It might be DiffusionGemma.

Not because it is the smartest model.

Because it points to something much bigger: dramatically faster AI from software, not just bigger hardware.

Most LLMs still generate text like a typewriter: one token after another.

DiffusionGemma takes a different path. It uses text diffusion to generate blocks of text in parallel, then refines them.

Google says this can deliver up to 4x faster text generation on GPUs.

That is not a small optimization.

In our early testing, the speedup was even more dramatic.

On a workload where a comparable Gemma 4 31B request took about 35 seconds, DiffusionGemma came back in about 5 seconds.

Roughly 7x faster in that case.

This is the kind of change that makes private AI feel different.

And the implication is easy to understate.

A 4x faster model is not just “the answer arrives sooner.”

It also means the same hardware can handle roughly 4x more requests.

At 7x, the same box starts to feel like a small cluster.

That changes latency, cost, capacity, and product design.

There are trade-offs.

DiffusionGemma is experimental. In our testing, it was not as reliable as standard Gemma 4 models on instruction following and structured response contracts.

So no, I would not treat it as the default private agent runner yet.

But that is not the point.

The point is the direction of travel.

My read: diffusion-style generation, parallel decoding, and related approaches are going to spread across model families.

If this becomes a 10x-class software improvement over time, it is not a bear case for AI hardware.

It is the opposite.

It blows open the floodgates.

Cheaper inference. Lower latency. More private deployment. More usage from the same infrastructure.

AI becomes much more accessible.

[Blair Milroy @blairmilroy]

@claudeai I’m glad to see Anthropic pushing dynamic workflows.

It validates a belief we’ve had at @clearideas for a long time:

The future of AI at work is not one giant prompt.

It’s governed, repeatable workflows over trusted data, with verification built in.

[Blair Milroy @blairmilroy]

@Clearideas I wanted people to be able to see how we think about the foundation: secure collaboration first, then AI workflows over trusted knowledge.

The Community Edition is smaller than the hosted product, but the philosophy is the same.

[Blair Milroy @blairmilroy]

Would love feedback from people building for small firms, self-hosting internal tools, or thinking seriously about AI governance before it becomes a mess.

More info: https://t.co/LtdkiBYapI

[Blair Milroy @blairmilroy]

I made part of Clear Ideas open source.

Why?

Because “just upload all your company knowledge into a black box and hope for the best” feels like a strategy we may all regret.

[Blair Milroy @blairmilroy]

The bigger idea:

AI over company knowledge needs a governed collaboration layer.

Files, access, workflows, context, citations, models — these things should not be scattered across random tools.

[Blair Milroy @blairmilroy]

@Clearideas Big progress for Gemini 3.5 Flash. Its on par with Pro, but the frontier gap is still wide.

[Blair Milroy @blairmilroy]

There’s a point in every “quick experiment” where you realize it has a README.

That point has been reached.

[Blair Milroy @blairmilroy]

@eleftherios_b Side project with a hamstrung codex apparently. @thsottiaux is resetting limits, so let's see what fast mode can push out this weekend.

[Blair Milroy @blairmilroy]

I did the classic founder thing where I said “this will be a quick little side project.”

Anyway, it’s almost ready.