All that would not help you with an AI training cluster interconnect. See Amin Vahdat's keynote at HotInterconnects 2025. Everyone is building a fabric for this stuff from scratch (Google/Falcon, Amazon/EFA, Azure/MANA, Cornelis/CN5000, and obviously Mellanox).
It is partly this and partly a funding vehicle for American next-gen computing. It is not that hard to estimate FP64 ballpark from a whole bunch of public statistics. And it takes a looot more than raw FLOPs to get a simulation working. And presumably a looot more to translate it into practice. And the openness makes it easier to talk to different vendors and not get in the way of them having all the H1Bs it takes to get these things to work.
Plus one think I like to say is that if a bullet is flying towards you, you could know everything about the chemistry of the gunpowder and the composition of the alloy without it affecting what happens next.
Today's limits are known and undisputable. Tomorrow's limits are a promise: some promises over-deliver, others under-deliver. :)
Regardless, to bring the discussion back to the claim at hand: at all points in future, we will need the ability to reason under partial information. "Absolutely flawlessly complete diagnostics" is an asymptotic goal we get closer to but never reach. This is both very doable for a disciplined human, and very hard to outsource completely to an LLM. Treated as tools operatored by competent users, they are magical. But they can not outperform their user.
> With alzheimer's an autopsy can tell for sure but that's not much help for a patient.
Ok let us unpack this statement.
For your point to hold, I would have to be saying "all kinds of practical diagnostics are invented now. No progress can be made in better diagnostics".
If Alzheimer's can be validated by slicing open a dead patient, there is a causal mechanical explanation for the disease. If we can not confirm that defect without slicing open the patient, that is a limitation of 2026 tools. The "One True Diagnosis" is an Oracle explanation that all real diagnostic techniques try to approach in the asymptotic sense, and it is helpful exactly because it clarifies in discussions like this.
There are going to be diseases where we do not yet have causal explanations. Or where we treat them without establishing them. Hypertension is one example: while technically it can be caused by vascular stiffness, some weirdness with the RAAS system, some hyperadrenergic weirdness, practically you get a lot of mileage out of just prescribing people telmisartan if they're old.
That does not mean the frontier of hypertension is settled, or the 10% who do not have a vascular stiffness problem would not benefit from better causal models of hypertension. Science is us continuously pushing back against the fog: of the tools we have in 2026, some are great, some are imperfect, some are promising etc.
Pyschiatry gets complicated because the failures are not mechanical. Even if you could image every single neuron in a person's head we do not have a very good way to define an algorithm for these issues. I do not have a good answer for psychiatry.
> This is a kind of thinking a lot of programmers fall prey to. The real world, outside of code, is a very fuzzy and inherently analog place.
Having said that, I would vehemently reject and push back against this, and without doubting your sincerity, characterize it as an ad hominem.
The vast majority of issues with the human body are mechanical in nature. Restricted blood flow, unwanted tissue, a broken bone, a bad valve etc. These are causal descriptions of "disease". Where causal descriptions exist, the "One True Diagnosis" principle holds. Psychiatry just happens to be unique in that it is a fuzzy science where we rely on checklists and ultimately all diagnosis is probabilistic.
EDIT:
> This is a kind of thinking a lot of programmers fall prey to. The real world, outside of code, is a very fuzzy and inherently analog place. There is very rarely one in any complex system having a complex problem needing a complex solution. At some point even the definition of diagnosis gets fuzzy.
I would also push back against this mindset in general. This is not a falsifiable claim, it is incoherence as an argument, and I do not need to be a programmer to hold this position.
That the real world is analog is irrelevant to its amenability to causal explanations. Or "fuzzy": "fuzzy" in this context just does not mean anything.
I am not trying to sound exasperated or win internet points, just impress this point on you and anyone reading this. We can write math to predict weather, make it tractable to solve using approximations, tolerate IEEE 754 weirdness, and finally tell what the clouds will do a week from now. This is nature telling us that there is a pattern to how it behaves, and it is the only weapon we have as scientists.
To say that nature is not amenable to explanations is a very defeatist thing to say: neither Newton nor Einstein nor any of the million-odd people that have built modern society would exist if nature did not have causal explanations. I urge you to reject this defeatist thinking.
> I think „the diagnosis” is over simplification and lots of professionals would disagree that there’s always a single one.
"The Diagnosis" does not mean "one root cause".
Situation: my car has some unexplained vibrations.
1. Mechanic A says that it is the engine mounts
2. Mechanic B says that it is some weirdness in how the exhaust assembly is hanging to the underbody
3. Mechanic C says that it is just my wife farting
I replace engine mounts and 40% of the problem is reduced. I then drive without my wife and the remaining 60% is solved.
"The Diagnosis" was: 40% mounts, 60% wife, 0% exhaust.
> There is no guarantee that the LLM will help you converge on anything.
Absolutely. The guarantee does not come from the LLM. The LLM is a simply an improved version of Google Search.
The guarantee can only come from a systemic application of epistemic discipline and reasoning, which is very much (smart) human territory.
Put it another way, I could make good decisions with/without LLMs, with some uncertain diagnostics as input. I would have to trawl through 50 papers myself, and it is possible that my decision arrives 5 years too late as a result. LLMs enable trawling and do some of the legwork in connecting the dots, but are ultimately only as capable as the orchestrating human.
Yeah I think the OP is muddling the point by conflating "physician's version of the diagnosis" with "The Diagnosis".
There is absolutely one "The Diagnosis". Human body is a machine, albeit a very complex one, and all measurement sources have noise. But they are all measuring one reality, and if there is a problem, there should be one explanation that all measurements align with. They can be noisy but can never be conflicting (instrument error notwithstanding).
Physicians' ability to arrive at "The Diagnosis" would vary, but it does not mean one does not exist. I am not sure if characterizing human body as derministic or not is relevant here.
Maybe I am missing something but I just find this wrong.
Everything is a puzzle: there is one "Truth" or one diagnosis. You (a smart human) should be able to converge on it by cross-examining your LLMs. By themselves, they have no interest in revealing this, no stakes, which makes them tools only useful at the hands of a capable investigator.
> Yes, they don't realize it or lie to themselves because ~50% dropout.
I think there's some misinterpretation here. Not staying on in academia after PhD (common/modal) is not the same as not getting to complete a PhD (rare).
In CS/tech, those who exit academia after PhDs get paid $300K-$500K in the industry. I don't think there's any misleading going on.
As someone who graduated with a 7.5 year long PhD last month,
I feel like PhD stipends are not a major problem. Like I got $40K in a low CoL area, but accounting for tuition and overheads I cost my advisor closer to $150K/year.
Now why are tuition and overheads that high is a reasonable question and it ties into inefficiencies in broader American administrative processes, but I cost society and taxpayers $150K/year, and that I'm doing it for societal benefit is honestly only partly true. The first 6 years was just me building real skills and letting myself be frustrated, and maybe in the last 1.5 years I did things that justify the $1M bill and more.
Even if I did eventually do things that justified the $1M bill, I think most students don't. The larger value IMO lies in a workforce trained in the failures and frustrations of grad school. While I could rattle of plenty of limitations of academia/grad school, I'm not entirely convinced that me being shortchanged/underpaid was one of those things.
That training is compute-bound and inference is memory-bound is well-known, but I don't think Nvidia deployments typically specialize for one vs the other.
One reason is that most clouds/neoclouds don't own workloads, and want fungibility. Given that you're spending a lot on H200s and what not it's good to also spend on the networking to make sure you can sell them to all kinds of customers. The Grok LPU in Vera Rubin is an inference-specific accelerator, and Cerebras is also inference-optimized so specialization is starting to happen.
MapReduce is nice but it doesn't, by itself, help you reason about pushdowns for one. Parquet, for example, can pushdown select/project/filter, and that's lost if you have MapReduce. And a reduce is just a shuffle + map, not very different from a distributed join. MapReduce as an escape hatch over what is fundamentally still relational algebra may be a good intuition.
Algebras are also nice for implementations. If you can decompose a domain into a few algebraic primitives you can write nice SIMD/CUDA kernels for those primitives.
To your point, I wonder if the 73 distinct transforms were just different defaults/usability wrappers over these. And you may also get into situations where kernels can be fused together or other batching constraints enable optimizations that nice algebraic primitives don't capture. But that's just systems---theory is useful in helping rethink API bloats and keeping us all honest.
Yes, GPT5-series thinking models are extremely pedantic and tedious. Any conversation with them is derailed because they start nitpicking something random.
But Codex/5.2 was substantially more effective than Claude at debugging complex C++ bugs until around Fall, when I was writing a lot more code.
I find Gemini 3 useless. It has regressed on hallucinations from Gemini 2.5, to the point where its output is no better than a random token stream despite all its benchmark outperformance. I would use Gemini 2.5 to help write papers and all, can't see to use Gemini 3 for anything. Gemini CLI also is very non-compliant and crazy.
While Arrow is amazing, it is only the C Data Interface that can be FFI'ed, which is pretty low level. If you have something higher-level like a table or a vector of recordbatches, you have to write quite a bit of FFI glue yourself. It is still performant because it's a tiny amount of metadata, but it can still be a bit tedious.
And the reason is ABI compatibility. Reasoning about ABI compatibility across different C++ versions and optimization levels and architectures can be a nightmare, let alone different programming languages.
The reason it works at all for Arrow is that the leaf levels of the data model are large contiguous columnar arrays, so reconstructing the higher layers still gets you a lot of value. The other domains where it works are tensors/DLPack and scientific arrays (Zarr etc). For arbitrary struct layouts across languages/architectures/versions, serdes is way more reliable than a universal ABI.
Hyperscalers do not need to achieve parity with Nvidia. There's a (let's say) 50% headroom in terms of profit margins, and plenty of headroom in terms of the complexity custom chip efforts need to implement: they don't need the complexity or generality of Nvidia's chips. If a simple architecture allows them to do inference at 50% of the TCO and 1/5th the complexity and reduce their Nvidia bill by 70% that's a solid win. I'm being fast and loose with numbers and Trainium clearly seems to have ambitions beyond inference, but given the hundreds of billions each cloud vendor is investing in the AI buildout, a couple billion on IP that you will own afterwards is a no brainer. Nvidia has good products and a solid head start but they're not unassailable or anything.
Yeah unfortunately no amount of manoeuvering is a substitute for a kill chain where a distributed web of sensors and relays and weapon carriers can result in an AAM being dispatched from any direction at lightspeed.
DoE compute budgets are ~10B USD across labs. AI training is a trillion-dollar workload. Different league.