Lance Balkus

@JonathanRoss321 Spark Decode Appliance is the assignment…. What do they say?

My job used to be writing code. Now it's marriage counseling for Codex and Claude Code.

Retweet if you’d like to see @JimmyKimmelLive fired. @ABC and @Disney are tagged.

@TaraBull Bad human.

Describe Jimmy Kimmel in one word

@basispointpod @stevenfiorillo The TAM is for SpaceX, not XAI. Starlink’s TAM is in the billion’s of people. They will be larger than Facebook.

The issues with the SpaceX IPO, as per @stevenfiorillo:

@JudgeDillard @BenSasse Senator Sasse sold food inside Memorial Stadium as a Senator. He’s such a great American. Humble, down to earth.

Every American should watch every second of this video. Thank you, @BenSasse.

Senator Ben Sasse, nearing the end of his life with the utmost grace and dignity, reflects on the foundations of patriotism. It begins with the devotion to your family, your neighbors, your faith, and your community. That’s the core idea of the American spirit - and the source of all the blessings America has today.

@TheAhmadOsman Gaming. I’ve heard the will still offer workstation, we shall see.

Addendum number 1

@jaketapper We Nebraskans are losing a thoughtful, intelligent human. His interview was incredible

Honestly Ben has won twitter and we should all just stop even trying

@TheRealJFeld @WallStreetApes “The auction only happened because Lone Star owned the note AFTER that process, and then ran a foreclosure process. Probably related to bond holder losses or something, so it wasn’t a typical deed in lieu.” People who are making money in RE deals aren’t losing property foreclos.

@BalkusLance @WallStreetApes Who is “they?”

A 20 story tower in San Francisco valued at $320 million dollars went up for auction There wasn’t a single bidder. Not even for $1 The property is now vacant, because no one wants to make the investment in Democrat run San Francisco High crime rates and homelessness are cited reasons for office buildings like this to not see purchasers Bidding for the property started at $80 million, but no offers. A massive decline in value WeWork used to be a major tenant occupying nearly 200,000 sq ft in this same building

@TheRealJFeld @WallStreetApes You miss the point. They went bankrupt. It wasn’t a value proposition. People lost millions of real dollars.

This is 100% wrong. Full stop. This deal actually had a pretty competitive process and got a ton of bids before it ever went to auction (I think it got hundreds of tours and maybe 60 bids). The auction only happened because Lone Star owned the note AFTER that process, and then ran a foreclosure process. Probably related to bond holder losses or something, so it wasn’t a typical deed in lieu. They came into that auction process with a big credit bid, which basically set the bar way higher than where most buyers had it underwritten. No one bid because it didn’t make sense at $216mm or whatever they set it at (they got it for $130mm), not because no one wanted it. It’s honestly a good asset. The suites are in solid shape, it’s generating income, and it’ll lease up over time. This kind of headline just ignores how these situations actually work.

@Jason This tax loophole needs closed. It has been misused for far too long long.

I’ve invested in 700 private companies Providing liquidity and loans to high performing founders is completely standard. In fact, investors encourage loans and secondary sales for founders to enable them to go for the Gold — which is exactly what SpaceX has done. The @NYTimesPR is presenting this as a controversy, when in fact it’s the private company playbook in 2026. Another hit job from the NYT

@aijoey @NousResearch @Teknium @NVIDIAAIDev @nvidia @ComfyUI Can’t wait to read about it.

my DGX Spark already runs 117 tok/s on heretic (NVFP4 35B-MoE) and 90 tok/s peak on aeon (NVFP4 27B-dense) out of the box without any optimization loop. their endpoint is roughly my starting point. however, im building the harness now and i will run it and see how much % improvement i get.

DGX Spark📽🔥Story Board gen from GPT Image v2, then brought into Comfy UI/Seedance 2.0 Story: Earth is facing total extinction. The tech guy activates “Hermes Agent” by @NousResearch , an advanced AI agent system. Hermes analyzes the global threat, deploys coordinated digital agents through satellites, energy grids, weather systems, autonomous drones, robotics, and communications networks. cc: @EHuanglu for workflow inspiration

@unusual_whales What @tedcruz needs to do before the election is free Fannie and Freddie. @BillAckman

Ted Cruz: "We ought to index capital gains to inflation. We need to give a real impact to the economy that will impact affordability and do so before election day."

@bscholl I would push back on a chunk of that. He shouldn’t be required to give back but ALL of the loopholes he uses need closed. Using stock as low collateral. So many ways to not pay taxes. This is why we eventually we will see a wealth tax.

The idea of billionaires needing to give back is deeply fucked up. Elon is rich because he gave the world many things it wouldn’t have had otherwise. The world pays him back merely with money. If you ask me, the world got the better end of the bargain.

@kaiostephens What was the prompt that kicked this off?

I ran Qwen3.6:27b to optimize itself in a recursive loop on my home server. Over 26 hours it went from 2.3tok/s to 84.3tok/s decode. It began on the home server, found there was no NVIDIA GPU, detected a CPU/RAM setup with 24 CPU threads and 93 GiB RAM, a 9060xt 16gb, then installed Hugging Face tooling remotely and started pulling GGUF quantizations. It benchmarked remote llama.cpp / llama-server runs across quantizations and flags: Found existing Qwen3.5-9B-Q8_0.gguf Downloaded / tested Qwen3.6-27B GGUF variants Compared Q6_K, Q5_K_M, Q4_K_M Ran server benchmarks over SSH against localhost:8080 Tested thread count, context, batch size, n_ubatch, --no-mmap, and memory-related flags Researched further speed paths: lower quantization, NUMA, huge pages, native CPU builds, cache/KV, TurboQuant, DFlash, speculative decoding, and automated tuning 1,524 tool calls 367 artifacts 345 memory addition 804 browser-control calls All of this from a model that can run on your computer. You don't need a better model. You need a better harness.

@KingKris90 @EventuallyWLTHY I’m not long or short but why wouldPalantir partner with a company if it was a zero?

@EventuallyWLTHY At what cost? Zero percent of their growth is organic

$ONDS backlog Last quarter: $68M Next quarter: $457M

@VaibhavSisinty I’d love to see the actual prompt that kicked this off to see what I’m doing wrong.

Codex is WILDDD!! Ran for 8 hours straight. No heartbeat. Gave it a brutal task: rewrite backend + data pipelines end to end (security, speed, edge functions, DB schemas) Came back… and it had actually made serious progress. This wasn’t a script. This felt like a 5x sr engineer grinding for 8 hours. Cost: 5,000+ credits 😅 I’ve seen 16-hour runs before, but this hit different. Feels like we’re moving from “AI tools” → “AI systems you manage” Shoutout to @gabrielchua and Vasundhara for unlocking me when I was stuck 🙌 What’s the longest you’ve let AI run on its own?

Every time I use Claude it seems to get worse at doing what I ask. Sonnet or Opus, doesn’t matter.