Custom Wetware

@realGeorgeHotz 1. Read The C Programming Language by K&R 2. Watch everything by Andrej Karpathy on youtube 3. Write code 4. Ask ChatGPT to explain stuff to you 5. Get familiar with debugging and profiling tools 6. Use git (do steps 3 to 6 while you do steps 1 and 2)

@unitcubed @levelsio I hope Tesla hires everyone to build its new fabs and euv machines

@levelsio Been working in ASML for almost a decade. I've started looking elsewhere just like hundreds of others and I am not even in the impacted group of people. Morale at an all time low. Any recruiters hmu, DMs are open.

Now the same thing that happened to Philips is happening to ASML It's being scavenged by midwit MBAs from McKinsey Terrible if true

@korzewarrior @notch The ++ are the problem

@notch Discord really is just irc++

Let's just go back to IRC again.

@ThePrimeagen Yuck, who would want to have sex with chatgpt? Qwen and Deepseek are much more attractive..

why is he hard while doing this?

@yacineMTB Imagine if they just signal boosted quality and suppressed everything else

"we are going to stop bots on x" Refresh the timeline 90% LLM generated anxiety bait from posters outside of the western world Sweet

@gfodor @fchollet The more unfamiliar a task is the more they have to rely on their context window to reason about it. Too much novelty at once will overload their ability to do that and they start making dumb mistakes.

@fchollet and yet every day I see evidence to the contrary using agents to deal with novel problems on unseen code bases.

This is more evidence that current frontier models remain completely reliant on content-level memorization, as opposed to higher-level generalizable knowledge (such as metalearning knowledge, problem-solving strategies...)

@SebAaltonen 5090 laptop is an undervolted 5080 desktop with bigger RAM chips

Important to realize: Nvidia RTX 5090 desktop version is pretty much 2x faster than RTX 5090 laptop. Apple is not yet close to Nvidia desktop/server. On desktop, the power consumption also matters less. And big fans reduce noise vs laptop.

M5 Max vs RTX 5090 laptop: - GPUs almost even when running on battery - Nvidia wins when charger connected - Apple 2x+ better battery life in heavy GPU workloads - Apple is much queter Apple CPU is clearly ahead of latest Intel in both MT + ST. youtu.be/CIhBCoQoymI?is…

@jukan05 @OmerCheeema Smells like an opportunity for Elon

@OmerCheeema What???

I live in Eindhoven, ASML town. Heard this from folks inside... Over the years ASML promoted a lot of strong EUV people into architect roles, group leads, management tracks. Built up serious layers. Classic growth pain. Then McKinsey comes in , says cut the management layers to speed things up. So now those same high performers, real good EUV experts, who got promoted are the ones on the block. About 3400 roles targeted, mostly management. Half will be reassigned, rest gone. Big hit in Veldhoven/Eindhoven area (~1400), some in US. Unncertainty is high, unions talking, details probably land around April. At the same time, ASML is still planning massive growth. The new campus near the airport just got final green light from city council. Construction starts soon, phased build-out. Long term they talk ~20,000 new jobs in the region (first wave ~5k by 2028). And the layoff packages are subpar. Philips is also based in the same town. And they had layoffs due to serious financial issues last year. Their layoff packages were much better than what ASML is offering. What a way to kill tje company culture. Especially at a time when the company is printing money.

@HaliReturns2026 @StealthQE4 Kamala would have been horrible in completely different ways. Hard to compare the two but this is what you get when both sides produce the worst candidates imaginable.

@StealthQE4 Blows my fucking mind how you all came to that conclusion and I’m terrified of your next awful decision

If we didn’t piss off Europe with the tariffs they’d probably be singing a different tune. Honestly this administration is a complete dumpster fire and I voted for Trump. There’s absolutely zero thought or analysis before they do anything. It’s all impulsive and ego.

@chamath Get rid of it before asteroid mining becomes economical

If I hypothetically owned an American resource with 5m oz of gold should I:

@LockePacem @TheAhmadOsman It can still run models that are big enough to be useful and the speed advantage is considerable. I would go for speed.

@CustomWetware @TheAhmadOsman Indeed. Though the benefit for inference on a DGX Spark with 128GB of unified memory is not to run a smaller model right? Or comparatively, 2x DGX Spark Cluster.

DGX Spark uses unified memory > 273 GB/s RTX PRO 6000 delivers > 1.8 TB/s (1792 GB/s) If someone told you they’re comparable, they’re wrong And this is exactly why llama.cpp isn’t the right tool here Try vLLM or SGLang on a GPU and you’ll see very different results

@CryptoElite007 @TheAhmadOsman Yeah and the RTX PRO 1.8 TB/s is actually only 900 GB/s, but the RTX PRO has more compute for prompt processing so it's still faster than the M3 Ultra

@TheAhmadOsman True. The Mac studio M3 ultra is around 870GB/s, waiting to see what the M5 ultra delivers... And I say this as someone that hates apple. I just want a solution that doesnt involve selling a house and stacking 8 cards...

@LockePacem @TheAhmadOsman A smaller model on a faster card will run much faster

What am I missing here? Even if the memory pool is smaller, the larger pool can run a larger model. I spend 15-minutes waiting for Opus-4.6 to complete relatively straight forward tasks, if qwen3.5-397b on 2x Sparks can do the same in a similar time-frame. Why is a Pro 6000 better?

@foley2k2 @TheAhmadOsman Obscenely expensive though, completely different price range than the RTX 6000

@TheAhmadOsman Isn't the B300 over 4 tb/sec? I hope you get a demo unit from Dell.

@jukan05 That's a lot of silicon to invest in the first generation of a chip that has never been used before

[EXCLUSIVE] Samsung Electronics Breaks Into OpenAI — Sole Supplier of 800 Million Gb HBM4 Samsung Electronics will become the first and sole supplier of next-generation High Bandwidth Memory 4 (HBM4) to OpenAI, the world’s largest artificial intelligence company. OpenAI plans to integrate Samsung’s HBM4 into its first-generation in-house AI chip, codenamed “Titan.” With this win following its earlier HBM4 supply agreement with NVIDIA, Samsung is being credited with cementing its leadership in the advanced AI chip market. According to industry sources on the 19th, Samsung Electronics has agreed to supply OpenAI with up to 800 million gigabits (Gb) of HBM4 (12-layer product) in the second half of this year. That volume represents approximately 7% of Samsung’s total planned HBM output for the year (over 11 billion Gb), and roughly 15% of its HBM4-specific production (approximately 5.5 billion Gb). The allocation is understood to be the third largest after NVIDIA and AMD. The HBM4 units Samsung will deliver are destined to sit alongside OpenAI’s first-ever AI chip, Titan Gen 1, developed in partnership with Broadcom. TSMC is expected to begin production in Q3, with a launch targeted for year-end. The deal is seen as particularly significant given that OpenAI — the company that ignited the generative AI boom with the launch of ChatGPT in 2022 — selected Samsung as its inaugural HBM supplier. OpenAI also sits at the center of the United States’ Stargate Project, a planned $500 billion AI infrastructure initiative. First Fruits After the JY Lee–Altman Meeting… AI Chip Orders Keep Coming OpenAI operates hundreds of thousands of AI chips across its data centers to deliver generative AI services, having relied heavily on NVIDIA’s general-purpose AI semiconductors. More recently, the company concluded that it needed its own custom chips optimized for inference workloads — a trend that has become central to next-generation AI development. An industry insider noted that “OpenAI has been investing heavily in R&D to successfully mass-produce its Titan chip,” adding that “Samsung satisfied the stringent HBM4 requirements that OpenAI set out, which is what made this deal possible.” Given that OpenAI has chosen Samsung as its first-ever HBM supplier, observers believe Samsung HBM is likely to be incorporated into future Titan generations as well. HBM stacks multiple DRAM dies vertically — similar to floors in an apartment building — delivering greater capacity and faster data transfer speeds than conventional DRAM, making it the memory of choice for AI applications. Micron Technology has projected that the global HBM market will grow from approximately $35 billion in revenue last year to around $100 billion by 2028. Samsung’s HBM business had a difficult stretch through last year, suffering back-to-back failures in NVIDIA’s qualification tests for HBM3 and HBM3E — a significant blow to the pride of the world’s top memory chipmaker. The turnaround came in May 2024, when Vice Chairman Jeon Young-hyun made the bold decision to redesign the DRAM at the core of Samsung’s HBM. This year, Samsung passed NVIDIA’s HBM4 qualification without a single design revision, enabling direct shipment of mass-production units. On March 18th, AMD announced it had designated Samsung as its preferred HBM4 supplier. Demand for the prior-generation HBM3E is also said to be surging. Samsung is reportedly targeting over 5 billion Gb of HBM3E for Google’s Tensor Processing Units (TPUs) — also developed in partnership with Broadcom — in addition to its NVIDIA supply commitments. Behind the OpenAI HBM4 deal, JY Lee, Chairman of Samsung Electronics, reportedly played a pivotal role. In October of last year, Lee met with OpenAI CEO Sam Altman and exchanged a Letter of Intent (LOI) covering the supply of cutting-edge AI memory including HBM, laying the groundwork for the agreement that has now come to fruition.

@ActAccordingly Memory is cyclical, we know. But the demand growth from AI is stronger and more persistent than previous cycles and there's a limit to how quickly supply can be increased. So the market will remain squeezed for longer despite high capex.

If you're wondering about the action in $MU today, I think you should look to Druckenmiller's sage advice on chemical stocks. If you don't look at the $MU print as anything but a commodity producer benefitting from a supply/demand imbalance you're probably not going to do well here. I don't make the rules, he does: "Chemical stocks, however, behave quite differently. In this industry, the key factor seems to be capacity. The ideal time to buy the chemical stocks is after a lot of capacity has left the industry and there’s a catalyst that you believe will trigger an increase in demand. Conversely, the ideal time to sell these stocks is when there are lots of announcements for new plants, not when the earnings turn down. The reason for this behavioral pattern is that expansion plans mean that earnings will go down in two to three years, and the stock market tends to anticipate such developments."

@sqs @kunchenxyz By making the product so slow that people only use it when their employer forces them to

@kunchenxyz Yeah. How does Atlassian handle this tension?

An uncomfortable truth about building agents/models: By default, your most lucrative, most-smitten customers will be those using intricate out-of-band techniques that are exorbitantly expensive and probably net negative (but that they love). It's a very weird incentive. You can't and don't want to indulge this. There's nothing wrong with experimentation, but if you saw what every agent company sees, you'd know this goes way beyond experimentation. Amp tries really hard to prevent this: limiting long context, showing prices, not recommending swarms or loops prematurely, strongly advising against big MCPs, killing features that have high usage but that aren't worth it anymore, and just generally staying away from any hype train we don't have a good gut feeling about. Pi and OpenCode are also particularly good and outspoken here. But if you have growth targets to hit, investors to pitch, and salespeople to keep happy, or if you didn't start this way from day 1, I can see it being tricky. At Amp, we're profitable, don't have salespeople, and have no sales/growth targets to hit, so we have it relatively easy. I often wonder what this tension is like inside other companies building agents. (And for the record: if you've shown me your Amp workflow and I haven't told you this directly, this post is not about you. :)

@asymcore @astridwilde1 Semiconductors are so high margin, they will be prioritized above lower margin industry if there's a shortage

@astridwilde1 Ok but doesn't the war affect this? How much can these go if Asia stops receiving adequate energy? The war is not stopping soon.

just getting started still underestimated memory is no longer cyclical. secular demand is going to outstrip supply through at least 2030 just own memory and take a nap for 3 years

@mattpocockuk @Stevie_658jjh The quality of context matters too, not just the quantity. They get distracted by their own low quality outputs.

@Stevie_658jjh Debugging a slow React Router loader. 200K was trying random stuff and getting confused. Cleared the context, and it solved it.

Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.

@francoisfleuret @capetorch 27B Q4 runs on 24GB RAM with decent context size, more if the context is also Q4

@capetorch Enough being?

1. What are the best open source coding / general purpose models? 2. What hardware to run them comfortably? 3. How do they compare to the flagships?

@kiaran_ritchie I wrote an opengl renderer first and then spent a day plus some debugging getting AI to translate it to vulkan.

This is nuts. Getting a Vulkan renderer off the ground is a Sisyphean rite of passage for graphics engineers.