cpldcpu

Isn't perplexity rather a "wrapper" company? Wouldn't acquiring a UX focused company bring enormous difficulties in integration, to retain Apples look and feel?

What seems to be missing for Apple is access to competitive foundation models. Perplexity has published a few finetuned models (https://openrouter.ai/provider/perplexity), but their focus does not seem to be own creating their own foundation models.

Furthermore, the entire angle on multimodality is also lacking, which is needed for true AI assistants.

Here you can see how i prompted it. I provided a similar paper (also generated with C. opus) as an example, but Opus took it from there:

https://claude.ai/share/963b66a7-930c-47a6-a4ea-d7e6993347fa

You can find the reference-paper also on Vixra: https://ai.vixra.org/abs/2506.0049

Sorry, this is just getting old...

Its a trite talking point and not the reason why there are so few consumer-AI companies in Europe.

But this is just the SFT - "distilled" model, not the one optimized with RL, right?

There is almost no information under the link

I think that's just a simplifying example. They would most likely not use an image sensor, but a photodetector with a broadband amplifier.

Is it ARM or x86 based?

I also strongly suspect that there are earlier sources.

However, IRAM looks like compute near memory where they will add an ALU to the memory chip. compute in memory is about using the memory array itself.

To be fair, CIM looked much less appealing before the advent of deep-learning with crazy vector lengths. So people rather tried to build something that allows more fine grained control of the operations.

Some more background information:

One of the original proposals for in-DRAM compute: https://users.ece.cmu.edu/~omutlu/pub/in-DRAM-bulk-AND-OR-ie...

First demonstration with off-the-shelf parts: https://parallel.princeton.edu/papers/micro19-gao.pdf

DRAM Bender, the tool they are using to implement this: https://github.com/CMU-SAFARI/DRAM-Bender

Memory-Centric Computing: Recent Advances in Processing-in-DRAMhttps://arxiv.org/abs/2412.19275

Berlin is not Germany.

Their approach seems very compelling, but I don't understand if/how they are building a differentiated product? The space of code agents is already pretty crowded.

That guy leads in with stating that he is missing "autocomplete" in claude code. Cleary a misunderstanding of the scope.

I thought now everything is about meritocracy? Have we been duped?

a goof part of team is actually located in europe

good point, it's certainly more of a trend with the younger generation and in the west.

But even living in a country that is perceived as "beer centric", i noticed that people are starting to be much more conscious of their alcohol consumption after covid. For the young generation, alcohol does not seem to play as big a role anymore and i would expect that this carries over to their work-life once they enter the work force.

Coming up next and already happening: Alcohol is phased out.

Thats not really correct.

It's actually the beginning of test time scaling. R1 has shown that a very simple reinforcement learning scheme can be used to teach the model how to think in a chain-of-though as an emergent property.

No addition pretraining data needed! Only more compute.

All this media frenzy around DS V1 makes me feel sick to my stomach.

It increased the noise in the AI space by orders of magnitude. Every media outlet is bombarding you with a relentless torrent of half-true information, exaggered interpretation of single facts and speculation.

Sometimes I wonder whether this is amplified by a state actor?

Also curious, how Minimax-01, which is also an excellent model with impressive improvements, went by completely unnocited.

The only good thing is that this certainly put an end to OpenAIs price gauging - pretty sure that $200/month individual plan is not the limit of their imagination.

The $6M that is thrown around is from the DS V3 paper and is for the cost of a single training run for DeepSeek V3 - the base model that R1 is built on.

The number does not include cost for personell, experiments, data preparation, chasing dead ends, and most importantly, it does not include the reinforcement learning step that made R1 good.

Furthermore, it is not factored in that both R3 and V1 are build on top of an enormous amount of synthetic data the was generated by other LLMs.

These benchmarks are mostly focused on math, which benefits a lot from an improved CoT and is also less sensitive to having "reduced knowledge" in smaller model.

Vibes are important in this case...

I think the main issue is that the average consumer does not know what to do with a raw transformer model.

While the base technology is now there and is rapidly improving, a lot of the "glue" and "plumbing" is still missing. What is the best way to integrate these tools into our normal workflows / daily lives and so on?

It will take time...

Articles like the one above are not very useful as they completely miss the big picture.

>n the first scenario, an investment in any AI model company that does not own its own compute is like buying a tar pit instead of an oil well. In other words, this future has AI like a commodity and the entity who wins is the one who can produce it at the lowest cost.

Why do you assume that AI will become a commodity that is only metered by access to compute?

Right now (since June 2024), Anthropic is ahead of the field in quality of their product, especially when it comes to programming. Even if O1/O3 beat them on benchmarks, they are still nowhere near when normalized for compute needs.

Can they sustain this? I don't know, but in the end this is very similar to known software or even SAAS business models. They are also in a somewhat synergetic relationship to Amazon.

Did office software ever become commoditized? Google and many more tried hard, but there is still the same company in the lead that was in the 90ies.

>LLMs are awesome but I haven't felt significant improvement since the original GP4 (only in speed).

Absolutely disagree. Are you using LLMs for coding? There has been a 10x (or whatever) improvement since GPT4.

I causally tracked the ability of LLMs to create a processore design in a HDL since 2023. I stopped in June of 2024, because Sonnet would basically oneshot the CPU, testbench and emulator. There are another substantional update of Sonnet in October 2024.

https://github.com/cpldcpu/LLM_HDL_Design

Voxel rendering is basically raymarching. Current GPUs can implement this easily as pixelshaders.

Plenty of examples on https://www.shadertoy.com/

yeah, the title is technically correct. but...

Which LLM is Amazon Q based on?

>That _should_ only make a difference for memory usage if your C compiler isn’t perfect

Considering that the PMC150 has an accumulator based 8 bit architecture which is almost hostile to C, it is safe to assume that the compiler is not perfect :)

Yes, you could implement it in a way where the first layer is streamed and accumulate on output activations in parallel in the memory. This would limit the memory requirements for the input activations, but would increase execution time, as more activiations have to be shuffled around.

In this case I am streaing from ROM anyways, so it does not matter if the inputs are read only once or multiple times.

There are no performance profiling mechanisms on these small devices, and the timers are rather coarse.

But it is easily possible to estimate the execute time:

- mulacc of one weight takes 11 clock cycles.

- There are 1696 weights in the model, each one is only touched once.

- We can assume ~25%-50% overhead for loops and housekeeping (1:4 unrolled)

=> ~23000-28000 clock cycles per inference, which is less than 2ms at 16MHz

Since this is an MLP, the inference time directly scales with the number of weights. (This would be different for a CNN)

As per veryfing on PMC150C - I considered using an LED for valid/nonvalid output. But iterating with OTP devices is quite tedious when you do not have an emulator. Since both devices are code compatible, we can assume that the code works on the smaller devices, though.

Yes, as far as i remember the limit was somewhere around 1kbyte total parameters size.