builder, ai, markets, and general data nerd | https://xitter.cf/t.co/zs1Hb6u5e0
In the lab 🧪
Joined March 2025
- Tweets19.1K
- Following1.3K
- Followers1.6K
- Likes17.5K
The Fed's cushion is thin.
Rate cuts are the standard recession fix, and with the Fed funds rate at 4 percent there's only about four points of room before zero. Krugman's argument is that an AI spending bust would hit harder than the late-90s IT bust, with less ammo to answer it.
Past zero, the remaining options are fiscal spending or the Fed buying assets, and the first one needs Congress to act.
“a bursting AI bubble would be a much bigger shock to the economy than the bursting IT bubble at the end of the 1990s that brought us to the edge of a liquidity trap. Yet the Fed has much less room to cut rates now than it did then: the current Fed funds rate is only 4 percent. So if huge AI spending is a bubble, and that bubble bursts, I can’t see how the U.S. can avoid plunging into a liquidity trap — a situation in which the Fed simply lacks the tools to get us out of a slump. If this should happen, the economy is likely to remain depressed for years.”
-@paulkrugman
Checking got cheap.
An agent solving 49 of 50 research-level physics problems at under $3 each pushes the cost somewhere else: confirming the problem it solved was the one you meant.
Math shows the size of that shift. OpenAI's August corpus had 20.6 MB of machine-checked proof against 55.6 KB of statements and definitions a human still has to read. That's 379 to 1, and I broke down what it means in the post below.
Those few pages are the only place a mistake can hide. Two days after the release, a Lean contributor passed 8 of the repo's challenges without doing any math.
dodatathings.dev/blog/verifi…
OpenAI just released 300 AI-written math proofs, each checked line by line by a computer. The repo still labels them unchecked, and that label is honest.
Two days after the release, a Lean contributor showed you could pass 8 of the repo's challenges without doing any math. The trick was rewriting the rule so nothing could ever meet it, which makes "nothing meets the rule" automatically true. The computer approved it, and it was right to, because it only checks the answer against the question it's handed.
AI made answers cheap and Lean made checks cheap. The question is still a person's job, and that's where this slipped. In OpenAI's August batch, machine-checked proof outweighed the statements a person still has to confirm by 379 to 1.
Tests on code, templates on contracts and validations on models carry the same stamp. So ask of each one: verified against what?
dodatathings.dev/blog/verifi…
Obsidian is just a folder.
Claude Code operates on the working directory. Launch it from a vault root and it has read/write access to every markdown file, plus CLAUDE.md and memory.md for persistent context. The agent never sees Obsidian, only files.
That's why "show every note that touches customer A" works as a plain-language prompt. The linking and search features were never what made the notes active.
Obsidian still earns its place as the phone reader and the pretty graph view. The agent layer doesn't need it, and I laid out the file-access mechanism in April.
dodatathings.dev/blog/claude…
Karpathy's LLM wiki pattern keeps showing up because it works. I've been running a version of it in production for 3 months inside Foliome, a personal finance agent that syncs every bank account, classifies transactions, and remembers my financial goals across sessions.
Same three operations. Ingest: the agent processes new bank data and updates its financial memory. Query: I ask how much we spent on the kids' activities this quarter and get an answer grounded in 5 years of classified transactions. Lint: periodic checks catch stale categorizations and flag anomalies.
I also built two Obsidian plugins using related patterns before the Claude Code + Obsidian wave hit X this month. The part nobody talks about yet is how far beyond note-taking this extends. Financial data. Research pipelines. Content systems. Anywhere you need an agent to accumulate context across sessions instead of rebuilding from scratch every time.
Wrote up everything I learned shipping these systems: dodatathings.dev/blog/claude…
Two economists, same zero.
Apollo's chief economist now says AI has delivered "zero payoff" to corporate America outside Big Tech. Goldman's chief economist said something similar about the macro numbers: AI added basically zero to 2025 US GDP growth.
The mechanism is accounting. About 75% of a data center's cost is imported parts, so GPU spending lands as imports and subtracts from GDP. I wrote up that math in April.
If both hold, Big Tech's own earnings are the only place the return on this capex can show up.
dodatathings.dev/blog/ai-cap…
The NBER study is real, the $600 billion number is real, and the 90 minutes per week stat is real. But the framing here misses what makes this genuinely scary.
Goldman Sachs Chief Economist Jan Hatzius told the Atlantic Council that AI investment contributed basically zero to US GDP growth in 2025. His reasoning: 75% of a data center cost is imported chips from Taiwan and South Korea. The spending registers as imports, which subtract from GDP. $650 billion flowing out of US balance sheets this year. Most of it landing in Taipei and Seoul.
The MIT study is worse. 95% of enterprise AI pilots showed no measurable P&L impact. But companies are accelerating spend, not slowing down. Amazon: $200B. Google: $175B. Microsoft: $150B. Meta: $115B. That exceeds the Apollo program and the US interstate highway system as a percentage of GDP.
The thing the NBER survey reveals that nobody wants to sit with: this is not a technology failure. 1.5 hours per week means the technology is available and people are barely touching it. The bottleneck is organizational, not technical. No data foundations, no measurable KPIs, no executive sponsorship beyond the all-hands announcement.
J.P. Morgan projects $5 trillion in total AI infrastructure through 2030. Somebody has to earn that back. So far, the productivity data says almost nobody is.
Broke down every number on both sides of this gap, from Goldman GDP calculation to the MIT pilot data to the builders actually generating ROI.
dodatathings.dev/blog/ai-cap…
Open weights are about to rank 4th.
Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, 3 points behind Claude Opus 5.5 at 58. It gains 11 points over K3, and the weights are due Nov 4.
I laid out the self-hosting math in June, when K2.6's 54 led the open table. The index has been re-based since, so that 54 no longer compares. The ~10M tokens/day breakeven still frames the decision, but now against a 3-point gap to the top closed model.
The model is 2.8T total parameters with 104B active. Once the weights publish, procurement teams can price real serving costs against that gap.
Kimi K2.6 just took #1 on the Artificial Analysis Index among open-weight models at a composite score of 54. That number sits inside frontier-closed territory from 6 months ago. The post-GLM, post-DeepSeek V4 ranking is now a Chinese open-weight at the top of the open table.
The AAI Index composite blends reasoning, coding, and instruction-following across 10 evals. At 54, K2.6 clears the 50 line that was Claude 3.5 Sonnet territory in mid-2025. Open weights, 1T parameters, $0.95 per million input tokens.
The practical read: every diligence team that wrote off open weights at the GPT-4 frontier has to re-run the procurement math at the GPT-4o frontier. Self-hosting flips above ~10M tokens/day, and K2.6 just moved the quality side of that calculation up a full tier.
The leaderboard refresh is a forcing function, not a vibe. The next 4 enterprise eval cycles will tell us whether the Chinese open stack is now the default fallback when the closed labs hike API pricing in Q3.
Source: artificialanalysis.ai/articl…
Open cyber capability is being dialed down.
Kimi K3.1 Pro leads open models on the Artificial Analysis Cyber Index at 64, ahead of MiMo's 56 and Mistral Large 4's 49.5. Moonshot says it stripped cyber training data from the open-weights version, which cut its own cyber benchmark solve rate from 90.5% to 81%.
For a security team that rules out Chinese origin, the shortlist stays the same. GPT-6 Astra refused 100% of CyberGym tasks and Opus 5.5 refused 98.5%, while Large 4 completed 81.7%. We laid out that provenance math in the post below.
The 64 could be a pre-stripping number. The weights teams can self-host on November 4th might land closer to Large 4 than the index implies.
dodatathings.dev/blog/procur…
Banks, insurers and hospitals can finally run a near-top AI model on their own servers that no Chinese lab built. It costs them about 7 points of quality.
Until now, regulated buyers had two bad options. Run Chinese weights and fight the compliance review, or drop a full tier down to Llama 4 or Gemma 4. Mistral Large 4 is the first close-to-top option in between.
It scores 38.4 on the Artificial Analysis Intelligence Index. The best Chinese models score 45 and 46.
Cost per task is messier. Large 4 runs $1.13, against $0.13 for MiMo and about $2.00 for GLM-5.3.
So the premium is real, and now you can price it. A team whose policy rules out Chinese weights pays it. A solo builder routes to whatever passes their eval cheapest.
dodatathings.dev/blog/procur…
Same sparsity, 11 more points.
Kimi K3.1 Pro keeps the 2.8T total / 104B active spec and scores 55 on the Artificial Analysis Intelligence Index, up from 44 for K3. That puts it 4th overall while still paying per-token compute for a 104B-class model.
I laid out the 16-of-896 expert routing in July. The gain came without growing the active slice.
Open weights are due November 4. AI token spend grew just 0.2% month over month in September, and cheaper near-frontier models like this one are part of the pricing pressure.
16 of 896 experts active per token on a 2.8T parameter model, only 104B actually firing. That ratio meansKimi K3 aims for frontier-level output while paying compute for a model 27x smaller than its total weight class.
That's the economics DeepSeek's MoE approach put on the map last year, pushed further.
Open weights at this sparsity level lower the cost of running near-frontier intelligence for anyone without a frontier-scale cluster.
arxiv.org/abs/2607.24653
Softmax silently rounds gradients to zero.
Inside softmax and cross-entropy, tiny gradients can round to exactly zero, or stay nonzero when they should vanish. That error can show up as sudden Slingshot loss spikes in training, and as distillation students drifting away from their teachers.
Higher precision is the obvious patch, and this fix skips it. On grokking benchmarks, test accuracy went from 94.3% to 100% in one setting. So far the evidence is only small-scale.
"A Fix for Softmax and Cross-Entropy"
Turns out two of the most fundamental functions in deep learning have numerical bugs that can actually destabilize training.
This paper proposes simple fixes to softmax and cross-entropy that prevent tiny gradients from incorrectly rounding to zero, or remaining nonzero when they should vanish, as these errors can cause sudden loss spikes and interfere with knowledge distillation.
On small-scale grokking benchmarks, the fixes eliminate Slingshot loss spikes and improve test accuracy from 94.3% to 100% in one setting. They also prevent student models from drifting away from their teachers during distillation, without requiring higher numerical precision.
alphaxiv.org/abs/2610.31642
Open weights are nearly caught up.
Once Kimi K3.1 Pro's weights ship, the open vs closed gap on the Artificial Analysis Intelligence Index drops from 12 points to 3 (55 vs Claude Opus 5.5 at 58).
Most of the jump is agentic work. Terminal-Bench 4.0 went from 13% to 67%, ahead of Claude Sonnet 5.5 at 64%. Agent workloads burn many model calls per task, so that's where API bills pile up.
Companies running agents at volume could start pricing a self-hosted model against their closed-lab invoices, with only 3 index points separating the two.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, 44). It trails only Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
Your best competitor is an employee.
Employees are adopting AI faster than their companies, often without telling management. AI-native startups now judge what AI can do very differently from incumbents.
I wrote up where that leads in March: a lawyer with 20 years of case law and Claude, a contractor who knows job sites. The people who know what their industry should look like already have the tool open.
The next wave of startups could come from inside incumbents, founded by their own senior staff.
Some might read this as "startups will rebuild your business using AI"
But this actually means domain experts finally have the tools to architect what their industry should be.
A lawyer with 20 years of case law now has Claude. A contractor who understands job sites has reasoning models. The window is for experts who can finally build.
That's the actual opportunity.