Tobias Mann

Aha if you set a system prompt “You are a helpful AI assistant” it will actually listen to you.

I’m currently fighting with OpenAI’s 20B param open model. It insists that Star Trek (1966) aired on CBS. And it refuses to admit it’s wrong.

Seems to be why the utilities are being so cautious about investment.

It ultimately comes down to the economics of energy at scale, right? Under the right circumstances nuclear could become economically attractive maybe?

I fear you’re right. Sure hope someone figure out fusion soon.

Hyundai to help build nuclear-powered datacenter in Texas : Fermi America will also have 4GW of gas generators for the Amarillo 'HyperGrid'

So real talk. Is this really a second renaissance for nuclear or a distraction from all the gas plants we’re building to fuel AI?

I want so badly for nuclear to make good on all the promises of my youth. And if can… space next?

www.theregister.com/2025/08/04/f...

Google agrees to pause AI workloads when power demand spikes : On hot summer days, air conditioning is rather more important than search summaries

The problem with AI isn’t just that it uses a lot of power, it’s how it uses it

Like a volume dial that jumps straight from 1 to 11 with nothing in between.

Modern grids weren’t designed for that. Which makes Googles latest move all the more important.

www.theregister.com/2025/08/04/g...

AMD preps rack-scale Helios systems to contend with Nvidia's Vera Rubin NVL144 House of Zen's biggest iron yet boasts 72 MI400 GPUs, 260 TBps of UALink bandwidth, and 2.9 exaFLOPS of FP4 AMD offered its best look yet at the rack-scale architecture that'll underpin its MI400-series GPUs in 2026 at its Advancing AI event in San Jose on Thursday.…

AMD preps rack-scale Helios systems to contend with Nvidia's Vera Rubin NVL144

AMD bets on rack-scale compute to boost AI efficiency 20x by 2030 Who'd have thunk? The bigger the iron, the more efficient it gets With Moore's Law on its last legs and datacenter power consumption a growing concern, AMD is embarking on an ambitious new goal to boost the energy efficiency of its chips 20-fold before 2030. And it sees rack-scale architectures as a key design point to get there.…

AMD bets on rack-scale compute to boost AI efficiency 20x by 2030

AMD zeroes in on Nvidia's Blackwell with MI350-series GPUs : And the House of Zen wants to put 128 of them in your rack

AMD's MI355X is a 1.4 kW liquid-cooled monster built to battle Nvidia's Blackwell

And the House of Zen wants to put 128 of them in your rack

www.theregister.com/2025/06/12/a...

#AdvancingAI #AMD

That’s the guy!

What again are we advancing? I can’t remember.

#AMD #AI

I have a lot of questions about this tech. It sounds neat, but:

Can it be scaled down to a pluggable?
Would that pluggable be reliable?
Do we even need it when the industry is moving toward CPO?
Will be cost effective? Existing pluggables are really really expensive like $4k+ a pop.

TDK claims optical breakthrough to tackle generative AI’s biggest bottleneck Response times of 20 trillionths of a second recorded in test to improve data transfer speeds

Just an FYI for those that don't know. AI needs lots of optics not because light is faster. After the electro-optical exchange at either end, its usually slower than copper. Optics are used because at 400G or 800G, copper can only go a few meters.

www.ft.com/content/f08d...

They’ve denied doing that. But they are doing sneaky stuff with LM Arena.

My guess is that they rushed Scout and Maverick out the door because it would have been embarrassing not to have a new Llama out before Llama Con.

If Llama-4.1 is a distill of the full Behemoth then it might be better.

I wanted Llama 4 to be good. It's not. It's possibly the worst model I've tried this year and that's including Gemma 3

Google offers 7th-gen Ironwood TPUs for AI, with AI-inspired comparisons Sure, we're doing FP8 versus a supercomputer's FP64. What of it? Cloud Next  Google's seventh-generation Tensor Processing Units (TPU), announced Wednesday, will soon be available to cloud customers to rent in pods of 256 or 9,216 chips.…

Google offers 7th-gen Ironwood TPUs for AI, with AI-inspired comparisons

Did Google seriously just claim its 9,216x TPUv7 pod had 24x the compute of the El Capitan supercomputer?

They do know FP8 is not FP64, right? Do they want @fclc.bsky.social coming after them?

Oh and the MI300A supports FP8! In an Apples to Apples fight, El Cap is still at least twice as fast.

Meta debuts first models from the Llama 4 herd : Says they’re done right as they don’t lean so far left

Eagerly awaiting 4-bit quants of Llama4 Scout to start testing at home.

In my the meantime @ssharwood.bsky.social has the low down on the launch over on @theregister.com www.theregister.com/2025/04/07/l...

#AI #Meta #Llama

Quick tip if you're trying to use #LLMs to automate tasks.

1. Tell the model you'd like help "writing a system prompt" to automate the task.

2. Coach it through the task, and have it update the prompt when it gets it right.

3. Test the prompt and iterate until its consistent.

#AI

Study suggests OpenAI isn't waiting for copyright exemption : GPT-4o likely trained on O’Reilly books without permission, figures appear to show

Fascinating research into the lengths major model builders, like OpenAI, are going to obtain data to train better models.

ICYMI: My latest for @theregister.com

www.theregister.com/2025/04/03/o...

So I guess we aren’t getting any product or roadmap updates. Maybe because not even the speakers know what’s “core” anymore.

Gaudi3’s TCO proposition sounds great compared to H100 / H200 but that’s not what it’s competing with anymore.
Customers cross shopping Gaudi3 will be looking at MI300X, MI325X, B200, and GB200 NVL72.

So……………

#IntelVision

Custom AMD chip powers Microsoft's new Azure HPC instances In-house DPU and HSM silicon also shown off

Only Intel can pull off things like custom Xeon 6s for AWS? That quirky All CPU die spin of MI300 AMD made for Microsoft, and ARM’s Compute Subsystem Based cloud CPUs beg to differ.

#IntelVision

www.theregister.com/2024/11/20/m...

Touting 18A’s gate all around and backside power delivery.
Process node has entered risk-production to scale up production for mass contract manufacturing.

More details coming at Foundry Direct Connect event.

#IntelVision

Moving on to Intel Foundry. I wonder if we’ll get any updates?

#IntelVision

Inflection AI's new offering ditches Nvidia for Intel Gaudi Struggling chipmaker scores another win

Now we’ve got Ted Shelton from Inflection AI talking about a collaboration announced in October last year.

Great win if they bought the chips, but they’re all rented in Intel’s Tiber cloud. Curious if/what they’re paying.

Find my write up on @theregister.com www.theregister.com/2024/10/07/i...

I’m getting the feeling that the take away from Intel Vision is “Our vision is whatever our customers tell us it is”

#IntelVision

Great another agentic RAG demo. Cool. What’s the point?
Intel does agentic RAG too? Congratulations! Join the club, it’s a big one.

Maybe tell me why I should take a risk on a dead end product like Gaudi3? Because that’s a tough sell.

You’d think at a conference called Intel Vision they’d talk more about where they’re going. So far its just about of execs looking in the rearview mirror patting themselves on the back.

#IntelVision