Todd Mostak

Truly awesome to see Jensen in his GTC keynote present @nvidia's joint partnership with @IBM to use cuDF to accelerate open source Velox and the Presto SQL engine in watsonx, showcasing how @Nestle achieves 5X faster queries for 83% less cost. Amazing collaboration between the Nvidia and IBM teams to make this a reality, and we're just getting started.

@firstadopter @wallstengine Basically that Rubin is looking to be much more cost competitive against TPUv8, perhaps retaking the lead.

@wallstengine Read what they wrote under the paywall re: TPU-Nvidia competitive balance

For those thinking that $GOOGL TPUs are miles behind $NVDA Blackwell, SemiAnalysis basically argues the opposite. They write that despite lower marketed peak FLOPs, TPUs can hit higher realized Model FLOP Utilization than Blackwell, so effective FLOPs tilt in Ironwood’s favor. In stress tests, Hopper only reached around 80% of peak, Blackwell landed in the 70s, and AMD’s MI300 series in the 50s to 60s, which is why the raw spec sheets overstate the real gap. There is a real debate now on how far behind $GOOGL TPUs actually are versus $NVDA GPUs, especially after the latest SemiAnalysis work on TPUv7 Ironwood vs Blackwell. On paper, Nvidia still advertises higher peak FLOPs, especially with GB200 and GB300, but once you haircut those headline numbers for what you can actually sustain under power and DVFS constraints, the distance shrinks a lot. TPUv7 closes most of the distance on raw throughput and memory versus GB200, then leans on TCO. From Google’s side the all in cost per chip at rack scale is meaningfully lower than a GB200 or GB300 system because they are not paying Nvidia’s full system margin. SemiAnalysis models that for a big, sophisticated user like Anthropic, TPUs can deliver roughly 30 to 50 % lower cost per effective PFLOP than GB300 if you invest in the compiler work and get solid MFU. Even if you only get around half the utilization of a GB300, you can still end up even or ahead on cost. The Anthropic deal is the proof point. Roughly 1 million TPUv7s tied to a 1 gigawatt plus buildout is not a side experiment, it is a top lab deciding TPUs are good enough on performance and better on economics for frontier training and high volume inference. SemiAnalysis also lays out how that threat already forced better pricing for others. OpenAI has not even deployed TPUs yet, but just having a credible alternative has helped them negotiate lower effective pricing on their Nvidia fleet.

@dedene @thsottiaux This ^

@thsottiaux A bash-mode, so you can execute something manually and the agent immediately has all the output as context.

Codex team is doing a weeklong hackathon starting next week, what should we build? Other than CLI auto-update, we're on that already

If it was already difficult to know what is real, we are quickly going to converge on a world where representation replaces reality. Generated with Sora 2 Pro.

Sora 2 Pro is really good. I asked it to recreate the scene of the 2010 Tahrir Square protests and it pulled together quite a convincing scene, complete with anti-Mubarak chants calling for the fall of the regime.

Sora 2 is just wild…

@springertimo @willdepue This.

@willdepue i feel like many many people on here need to stop adjusting their beliefs after every launch; it's actually crazy how much of a rollercoaster most people's beliefs are.

i feel like nobody believes in AGI anymore

@Alibaba_Qwen are you guys still planning to release Qwen3-32B-Base? It seems to be the only Qwen 3 model variant aside from the 235B version without a base model.

@JagersbergKnut @Dan_Jeffries1 I think there’s an argument that China “wastes” far more resources with easy money for whatever the Party’s current goals are (ie electric car or chip production), but out of the waste emerges an actual edge, if just by the law of large numbers.

@Dan_Jeffries1 while I share the techno-optimist attitude, and I think one should try to extract whatever innovative capacity can be mustered to tackle the challenges, we do have finite time and resources to make a difference. China is more focussed than we are and your VCs also waste money

Two books I just read with essentially the same message: We've forgotten how to build in the west and a poisonous mentality and politics of scarcity and degrowth has taken hold on both the left and the right. For the left that scarcity manifests as we've got to use less, slow down everything from consumption to energy use. For the right, it manifests as immigrants are stealing all the jobs and houses and we've got to limit access to them. Both of these are dead end philosophies that have never made any society richer or more prosperous. The answer is the build. We won't deflate our way out of an energy or housing crisis. We won't get there by closing every door. We've got to build. More energy. More houses. More factories. Build. Build. Build. Nuclear. Solar. Wind. All of it. Not some of it. All of it. Now. Degrowth will lead to shortages and riots and more political extremism. And eventually war. That is history in a nut shell. In times of plenty we work together and grow. In times of scarcity we kill each other. If we don't have enough homes we've got to build more homes. Simple as that. To do that we've got to slash through a thicket of red tape in zoning laws and absurdly long permit procedures as builders fill out endless forms and submit environmental impact assessments that do nothing to protect the environment and everything to make life miserable for the builder and the young couple trying to buy their first apartment that is now out of reach. We're at a tipping point now. The politics of scarcity is winning. It's taken hold like a sickening rot that's poisoning minds and policies and setting us down a dark path. We still have a chance to turn it around before that tipping point is crossed but when the Rubicon is crossed it will play out like a vicious storm, and then there is no stopping it and its dark energy will rip apart everything in its path before it finally blows itself out in exhaustion and subsides again, leaving a trail of tears behind it. Nobody wants the storm except the insane and the short sighted and the people who make their living spreading rage and fear. But it's coming if we don't begin to think differenty and build new things and trust in science and technology to make a better tomorrow. Build.

We recently expanded our support for the Uber H3 hierarchical geospatial index to HeavyDB, allowing users to easily aggregate, visualize, and join data using the H3 grid. Check out our new blog to learn more and see examples of this powerful new capability in action.

Enabling real-time geospatial joins at scale has always been a key goal of ours at @heavy_ai. Check out our blog benchmarking geo join performance in GPU-accelerated HeavyDB against Postgres, DuckDB, Snowflake, BigQuery, and Redshift.

Post here: heavy.ai/blog/connect-t…

@BelugaClyde @Fandango Same, bought two tickets for my sons and one for myself expecting they would both get Jetpack promo codes. Now we only have one code. Would have bought separately if we had known. @Fandango if you care about your loyal customers you need to make this right.

@Fandango please tell me this ridiculous scenario is incorrect. I bought 11 tickets for the new minecraft movie specifically on your platform due to the offer for the in-game jet pack. After the purchase i received only one redemption code for one in-game jetpack.

@elitedoorworks @musk_news13 @grok It could also be 25 I suppose but that breaks the symmetry of all numbers on the top half of the circle multiples by 5 equaling their counterpart on the bottom half.

@elitedoorworks @musk_news13 @grok It’s 1. For each pair of numbers on the opposite sides of the circle, the smaller number multiplied by 5 equals the larger number. Here ?X5=5 so it’s 1.

@USGS in conjunction with @UCSD recently released LiDAR-derived Digital Elevation Model (DEM) and Digital Surface Model (DSM) datasets for the areas impacted by the devastating Palisades and Eaton fires. First we ingested each dataset into @heavy_ai (roughly 1.2 billion records each), and then pulled @OvertureMaps building footprint data for the area. We then joined the Overture building footprints to the the DEM data (essentially ground elevation) and DSM data (which includes the height of man-made structures) to get a map of which structures were destroyed by the fire, generally the purple structures on the map below, which are computed to have near-zero structure height post-fire. The entire query, with two 1B+ row geospatial joins between the elevation and buildings datasets, takes less than 100ms on an @nvidia Grace Hopper GH200 GPU on @Vultr cloud, meaning the outputs can be interactively visualized. We are currently diving deeper into the data, and augmenting it with imagery to try to build a burn risk model based on the location of vegetation relative to structures. Our hope is that approaches like this that can fuse multiple geospatial datasets can help mitigate risks of future wildfires.

We're proud to announce a new partnership with @Ookla, a leader in network intelligence. With Ookla's wide array of unique crowdsourced and drive test data, paired with @heavy_ai, organizations can optimize, plan, and pinpoint issues in their networks faster than ever before.

Interactively explore over 20 billion records of ship AIS data in our new demo of HEAVY.AI on the @nvidia Grace Hopper Superchip, running in @Vultr Cloud. See our blog for more info: heavy.ai/blog/interacti…

@zied_houidi Wonder if you could programmatically generate the scenarios but then have an LLM turn them into word problems to add some entropy to the prompts?

@ToddMostak We're curious exactly about this too! ;) Let it be RL or SFT; could better episodic recall lead to better reasoning?

1/12 We just found something unsettling: Today's most advanced AI models - including the latest powerhouse reasoning models - can't keep track of what actually happened. Even in a simple conversation. Our ICLR'25 paper reveals why this matters 🧵