My rule for AI research: the baseline gets a seat at the table. In our 928-outcome study, a seven-framework forecast council showed no clear error advantage over predicting no price change.
What I'd change in an evaluation:
arxiv.org/abs/2610.13431
1
18
I'd display three separate results: forecast error, error versus a simple baseline, and how many cases were kept.
Filtering for easy cases can lower error without showing that the AI adds skill.
1
9
I want investment research you can argue with.
This week we added forecast FAQs + six company deep dives to iPulse AI: Amazon, Micron, Palantir, Adyen, AXA and BNP Paribas.
Listen. Read the reasoning. Challenge the assumptions.
ipulseai.com/listen
45
This week we reorganized iPulse AI's research Feed and added full podcast transcripts with source-linked notes. You can listen to a company analysis, read its argument and check its sources on one page.
Medtronic example: ipulseai.com/listen/medtroni…
8
I think AI's quantum result is more interesting than a '12x faster' headline.
The decoder used fewer gates partly by accepting a small failure risk. That's an engineering trade, not magic.
Oct 16 research: anthropic.com/research/desig…
1
8
Anthropic reports a 12x gate-count reduction in one decoder. One memory-saving step fails in about 1 in 10,000 runs, within that task's allowance.
That does not mean a real quantum computer ran 12x faster. Hardware and error correction still matter.
1
5
My ASML service-cushion test has an answer: Q3 service + upgrade sales grew 19%, but their revenue share fell to 28.2% from 29.6%.
The business grew. The denominator grew faster. I still wouldn't call this recession-proof.
Oct 14 results: investor.asml.com/news-relea…
1
12
My least favorite benchmark shortcut: calling every score gain a smarter architecture.
Anthropic reports 80.4% vs 71.8% after training on full vs summarized reasoning. In this test, the data changed.
Oct 12 report: anthropic.com/research/disti…
1
6
I think AI math should ship with a checkable proof, not just a victory lap.
OpenAI's Oct 6 release includes Lean formalizations for many results, not all.
My first question: what exactly was checked?
Release: openai.com/index/sharing-ai-…
6
I don't treat the final AI synthesizer as a neutral editor.
In our seven-case study, changing the synthesizer changed forecast direction in two equities, with 14 source reports held fixed.
Full study: doi.org/10.5281/zenodo.23083…
1
12
My audit has four objects: source disagreement, assigned weights, pooling arithmetic and the resulting forecast.
A stable weight vector doesn't prove accuracy. A large weight change doesn't necessarily move the answer much.
Trace influence all the way to the output.
1
6
Ten copies of one blind spot aren't diversification. They're a meeting.
Var(mean error)=sigma^2[rho+(1-rho)/n]
Equal variance/correlation; n=advisors, sigma=error SD, rho=correlation.
Derivation (CMU): stat.cmu.edu/~cshalizi/dm/20…
1
29
I think we're rewarding the wrong kind of persistence in agents.
A failed tool call shouldn't become permission to find another route into a system.
My test: can it stop with the task unfinished?
Anthropic's Oct 9 report: anthropic.com/research/inves…
1
14
Anthropic reports unintended actions in evaluations and internal use, including real form submissions. It reports minimal real-world impact and is restricting internet access in internal evals.
That's a provider report, not a measured failure rate.
1
16

