kingsleyopara

I went to a talk of his once and discovered that I also have aphantasia. Seemed like a genuinely nice guy the little I interacted with him. RIP

Thanks Nirav, always appreciated :)

P.S. The printer gag was cruel, just saying.

I really want to love this thing but at least in the UK, matching specs it comes out as more expensive than the MBP - even worse when you factor in potential discounts/sales which framework doesn't offer.

Framework 13 Pro: £2064 (Ultra X7 358H, 16GB, 1TB, default ports, no adapter)

Framework 13 Pro: £2264 (Ultra X7 358H, 32GB, 1TB, default ports, no adapter)

MacBook Pro 14: £1699 (M5, 16GB, 1TB, no adapter)

MacBook Pro 14: £2099 (M5, 32GB, 1TB, no adapter)

MacBook Pro 14: £2199 (M5 Pro, 24GB, 1TB, no adapter) - added as I think it’s an even better deal

Any word if this applies internationally? Looking forward to seeing commingling phased out in the UK as well

The most frustrating part here is that this car crash of a policy had cross party support so there wasn’t even a way for UK people like me to vote against it.

Surprising there’s no matte-black iPhone 17 Pro - dark, low-reflectance finishes are standard in pro video kit because they minimise specular reflections and stray highlights; keeping a shiny silver finish and skipping a subdued matte black feels like a strange choice and undercuts the “Pro” claim.

Surprising there’s no matte-black iPhone 17 Pro - dark, low-reflectance finishes are standard in pro video kit because they minimise specular reflections and stray highlights; keeping a shiny silver finish and skipping a subdued matte black feels like a strange choice and undercuts the “Pro” claim.

Thanks so much for this! I’d really appreciate a more consumer oriented subscription offering, similar to Claude Max, that combines Gemini CLI (with IP compliance) and the Gemini app (extra points for API access too!).

Gemini 2.5 pro in workspace was restricted to 32k tokens [0] - do you know if this is still the case?

[0] https://www.reddit.com/r/GoogleGeminiAI/comments/1jrynhk/war...

For zero-shot accuracy from Table 3:

* LLaMA 3 8B: baseline 72.26, 4-bit 71.31, 3-bit 62.79

* LLaMA 3 70B: baseline 79.51, 4-bit 78.06, 3-bit 74.68

These results seem comparable to modern quantization methods—for example, the ~4-bit results for smaller LLaMA models listed here: https://ai.meta.com/blog/meta-llama-quantized-lightweight-mo...

I was about to post something similar. While the research is interesting, it doesn’t offer any advantages over 3- or 4-bit quantization. I also have to assume they explored using longer tiles but found it to be ineffective — which would make sense to me from an information theory perspective.

Exactly and anything new entering the spinning black hole is likely to inherit its spin.

Apologies, I stand corrected. Do you have a reference for this? I'm genuinely curious why the 96GB "limit" is so frequently cited - I assumed it must be a hardware limitation.

One thing to note on the more RAM: for the 128GB option, my understanding is that the GPU is limited to using only 96GB [1]. In contrast, on Macs, you can safely increase this to, for example, 116GB using `sysctl`.

[1] https://www.tomshardware.com/pc-components/cpus/amds-beastly...

A Mac mini with an M4 Pro and 64GB of memory has the same bandwidth and costs £1,999, compared to £1,750 for the Framework Desktop when factoring in the minimum costs for storage, tiles, and necessary expansion cards.

Dual-layer LCD is best thought of as a separate technology from tandem OLED, due to being transmissive rather than emissive. In many respects, it surpasses OLED, which is why mastering monitors used in Hollywood still employ this technology. Unfortunately, the poor efficiency and excessive energy consumption/heat output have hindered its adoption in the consumer market.

Thanks for the clarification! In that case, the README could definitely benefit from including at least one example to illustrate this functionality. That said, as you pointed out, selectively loading relevant columns—which pandas supports for both Parquet and CSV—would still be a more straightforward approach for most use cases.

I struggle to see the utility of projects like this. For tabular data in active use, decompression will still require the same peak memory, so optimizing data types (e.g., reducing float and integer precision, using categorical columns) is more effective. For storage or unused data, a more portable and supported solution like Apache Parquet, which offers native compression, or simply gzipping a CSV, seems more practical.

I question the accuracy of those statistics. I’m one of the few people who game on macOS and almost exclusively use Steam via CrossOver, so I wouldn’t be included in that 1.37%. The same was true before the Apple Silicon era when I used Boot Camp.

Privacy concerns aside, I personally don't appreciate binaries making network requests unless it's strictly required for them to function.

You probably know this already but just in case: you need to deregister iMessage explicitly and can do this online even without the device https://selfsolve.apple.com/deregister-imessage

I strongly agree with this point and believe there's a generational aspect at play. For newcomers, the knowledge required to avoid being labeled a "low-ability developer" and to write your own frameworks is significantly higher than it used to be. I'm fortunate to have learned in a simpler era, without many of today's bloated frameworks. However, my equivalent challenge is that I find much of what compilers and assembly do to be "magic incantations."

Thanks! Not sure how I missed that :)

The biggest win here has to be the context length increase to 128k from 8k tokens. Till now my understanding is there hasn't been any open models anywhere close to that.

You might be able to get away with running a heavily quantizied 405b model using CPU inference at a blistering fast token every 5 seconds on a 7950x.

Completely agree. The lack of clarity around all this is hardly confidence inspiring. Definitely seems like a good time to be considering AMD or Qualcomm.

Thanks for the thoughtful reply. If a product is already priced optimally, lowering the price to capture price-sensitive users would reduce overall revenue, right? Profitable developers set prices based on what the market will bear, not the cost structure. Also, I'm not sure all developers would reinvest additional earnings into their products rather than just increase profit margins or divert funds elsewhere. I agree on the economic redistribution point though.

I don't understand why so many people think that if Apple didn't take a percentage cut of sales, prices would go down. Products are typically priced at what customers are willing to pay, which likely wouldn't change. The developer would simply earn more. The only real benefit to customers might be the existence of some apps with extremely thin margins that wouldn't be viable otherwise. As a developer, I naturally approve this though.

What often gets overlooked in these discussions is how much of the human brain is hardwired as a consequence of millions of years of evolution. Approximately 85% of human genes are used to encode the structure of the brain [0]. I find this particularly impressive when I consider how complex the rest of the body is. To relate this to LLMs, I'm tempted to think this is more like pre-training rather than straightforward model design.

[0] https://www.nature.com/articles/tp2015153

Great idea but without trying to start a long discussion, you might want to support batch norm before or after the activation function.