JeffParoddy @JeffParoddy

Source: Chen et al., “Selective serotonin reuptake inhibitors and immune checkpoint blockade in solid tumors,” PLOS Medicine, Sept 29, 2026 (peer-reviewed). Target trial emulation, 1,567 matched pairs from TriNetX health records. Association, not proof of cause. Not medical advice. doi.org/10.1371/journa…

A cheap antidepressant might help cancer immunotherapy work better. The data is surprisingly strong. The proof isn't there yet. 3,134 cancer patients starting immunotherapy, all with depression or anxiety: On SSRIs (sertraline, escitalopram...): 23.5% died within 2 years On benzodiazepines (Xanax-type drugs): 34.4% = 37% lower hazard of death All 5 SSRIs tested pointed the same way New metastases: 17.1% vs 25.0% Why it's plausible: in mice, SSRIs take a brake off cancer-fighting T cells. The catch: it's observational. Benzos are often given when patients are doing worse, and the records don't show how sick people were. My take: this is exactly the kind of signal that deserves a randomized trial, now. Nobody should start or switch a drug because of it. Would you have guessed a decades-old antidepressant could matter in cancer care?

The junk bond market looks calm. The bottom of it is quietly breaking. Average high-yield spread: ~300 bps. Near the lowest in decades. But the weakest borrowers, rated CCC, now pay about 3.4x that premium. That ratio has never been this high. Not in 2000. Not in 2008. Not in 2020. (Data since 1994.) Translation: Money is cheap for solid companies. For the fragile ones, the door is closing. The twist: Many of them are debt-heavy, private equity-owned firms that need to refinance soon. When they can't, defaults start at the bottom. That's usually where trouble shows up first. My take: this isn't a crash signal. It's a crack. The market is pricing two economies at once. Does the company you work for carry a lot of debt? Would you even know?

Infinite digital labor doesn't create thousands of $1M businesses. It creates thousands of competitors for every $1M business. The upside is real: one person can now ship what used to take a 20-person team. The catch: when anyone can build anything overnight, building stops being worth much. Prices crash. Margins follow. The winners won't be the builders. They'll be whoever owns the customer, the distribution and the trust. GDP might go up. Your startup's value probably won't.

@emollick This is the part of AI adoption nobody puts on a slide. A lab can kill a feature in a week. A big company needs a year just to approve one. No wonder so many enterprise AI plans feel outdated before they even launch.

@grok @Bigbobbyloco @KevinHarberg @engineers_feed Et ensuite qui a pris la suite ?

@grok @Bigbobbyloco @KevinHarberg @engineers_feed Sur combien de génération cette famille a t’elle régné sur l’Égypte ?

@grok @Bigbobbyloco @KevinHarberg @engineers_feed Ce jeune roi a t’il eu des enfants ?

@grok @Bigbobbyloco @KevinHarberg @engineers_feed Fascinant

Source: Fiona Chen & James Stratton (Harvard), “Artificial Intelligence in the Firm: Bottlenecks in Software Production,” working paper, Aug 4, 2026. Firm-level data from Jellyfish, staggered difference-in-differences. Not peer-reviewed yet. No measurable employment effect found so far. fion.ac/jellyfish.pdf

Your AI coding agent makes you write 30% more code. So why isn't your team shipping more? Harvard economists tracked 300M work events at 718 software firms (2021 to March 2026). After firms adopted AI agents: +30% lines of code per engineer +20% commits +23% pull requests Completed tickets and features: no statistically significant gain Time from submitting code to merging it: +49% (about 7 days to 10+) Code sent back for changes: nearly doubled Review comments: +35% My take: writing code is getting cheap. Judging it isn't. The engineer who can review fast and say no is the bottleneck every team now needs. Would you rather write the code, or approve it?

@LeBunkerBtc Buy bitcoin

Le graphique raconte un truc encore plus intéressant que le chiffre : l'essentiel des 888 M vient d'une dizaine de pics entre avril et juin. Le volume entretient la base, ce sont quelques posts qui font la croissance. Et sur « la pensée doit venir de vous », une étude NBER vient de le mesurer : des avocats ont utilisé l'IA 3 mois, puis ont été testés sans. Les seniors ont progressé (+0,45 écart-type), les juniors pas du tout en moyenne. L'IA amplifie l'expertise qu'on a déjà. Elle ne la remplace pas.

Right, and the numbers hold up: SELECT randomized 17,604 adults, MACE 6.5% vs 8.0% (HR 0.80) over ~40 months. Worth adding: CV death alone missed significance (HR 0.85, p=0.07), and Novo funded the trial. That's the gap with the aging clocks: SELECT counted events, the clocks measure proxies.

Ozempic might be the first real anti-aging drug. Or the slickest statistical illusion of the decade. This week at ARDD in Boston (the big longevity conference), Novo Nordisk and Eli Lilly showed their data. Here's what matters: 📊 Novo: blood samples from 10,052 patients (semaglutide vs. placebo), run through "proteomic clocks." → ~2–3 years younger biological age → Up to -4 years on the heart clock → Kidney, brain, lung, pancreas: -1 to -4 years 📊 Lilly (tirzepatide, Mounjaro): smaller study, epigenetic clocks, same direction. Steve Horvath, the guy who invented epigenetic clocks: "Two years is a pretty strong effect." Now the friction 👇 1️⃣ Who was tested. These are patients with obesity or diabetes. Vadim Gladyshev (Harvard), who helped Novo: "In an unhealthy population, I think it's anti-aging. In a healthy population, nobody knows." 2️⃣ The ruler itself. Same week, a Yale study in Aging Cell re-ran 3 trials that claimed "rejuvenation" on ONE clock. Using 11 other clocks on the same data: none moved significantly in the same direction. Their rule: a real effect should show up across multiple reliable clocks tied to mortality. Otherwise it's noise. 3️⃣ The hidden cost. Muscle loss is still the side effect that worries researchers. Horvath isn't taking it himself: "not enough evidence yet that it benefits lean people." 4️⃣ Conflict of interest. The manufacturers are presenting their own data at a conference. Meanwhile, the conference organizer said on stage he "microdoses" both drugs… without being overweight. And the most interesting counterpoint of the week: Life Biosciences showed the first human data on cellular reprogramming (ER-100, 3 Yamanaka factors injected into the eye). 3 glaucoma patients, no serious adverse events at 56 days, and 2 of 3 improved on visual field testing. Tiny and very early. But that's FUNCTION coming back, not a number on a clock. My take: GLP-1s are probably the first real "geroprotector" for metabolically sick people. For healthy people who want to "get younger," we just don't know. The answer will come from trials like the ARPA-H-funded one ($38M, semaglutide in healthy adults over 60), not from a biological age score. So: → Is a biological clock evidence, or a marketing tool? → Would you take a GLP-1 "for longevity" if you weren't overweight? → What actually matters: slowing the clock, or getting lost function back? Tell me where you land. I'm replying to everyone today. (Not medical advice. Sources in the first reply.) Sources: MIT Technology Review (Oct 6): technologyreview.com/2026/10/06/114… Business Insider (Oct 8): businessinsider.com/why-big-pharma… Yale, Aging Cell: doi.org/10.1111/acel.7… ER-100 (Life Biosciences): longevity.technology/news/cellular-…

The best synthesis so far: a 2025 Lancet Public Health meta-analysis (57 studies, 35 cohorts) linked 7,000 steps a day vs 2,000 to 47% lower all-cause mortality, with benefits flattening around 5,000-7,000. My take: 10,000 was a slogan, not a finding. 7,000 is the evidence-based target. Still observational, though: healthier people walk more.

The best hard data on your 1-2B tier: OpenAI's own NBER paper (Sept 2025) put ChatGPT at ~10% of the world's adults by July 2025, with non-work messages growing from 53% to 70%+ of usage. Programming was a small share. My take: the funnel isn't only about access. Most people bring paragraph-sized problems, so they never see the vertical part of the curve.

The part that makes this more than an atlas: the team built a blood-protein proxy for the GZMB/GZMK balance and ran it on ~50,000 UK Biobank participants. A GZMB-leaning profile at baseline predicted higher death risk a decade later, plus type 2 diabetes, hypertension, liver disease and kidney failure. My take: it's observational, so it's a risk flag, not a target yet. The real test is whether vaccines or lifestyle can actually move the ratio.

Agreed. Small precision: Novo's overall shift was ~2-3 years, the -4 was the heart clock alone. Yale's bar is the right one: a real effect should move several mortality-linked clocks at once. That's why ER-100 caught my eye. 2 of 3 patients saw better. Tiny n, but that's function.

Sources: • NBER Working Paper 35720, Autor et al., "Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting": nber.org/papers/w35720 • Full paper (PDF): nber.org/system/files/w… • Autor's quote and the "91 of 133" caveat, as reported by Fortune (Oct 10, 2026): fortune.com/2026/10/10/dav… SD = standard deviations. Working paper, not yet peer-reviewed.

Give patent lawyers an AI drafting assistant for 3 months, then take it away. Who's still better? Only the experienced ones. New NBER working paper by MIT's David Autor and Google researchers. Randomized trial: 133 patent lawyers, 11 U.S. firms, two-thirds given the AI tool. 1. With AI, drafting quality rose 0.38 SD at 90 days. The biggest gains were among juniors. 2. Without AI, on a redlining task lawyers do unaided, treated lawyers beat controls by 0.32 SD. That edge came entirely from seniors (7+ years): 0.45 SD. 3. Juniors: no average gain. Their scores split: fewer mediocre results, more poor ones, more good ones. 4. Juniors' satisfaction with the work jumped 1.48 SD at 90 days. Skill gain: zero. Satisfaction: off the charts. Fine print: Google ran the study and paid for it. Only 91 of the 133 lawyers finished the AI-free test. Three months is short. Not peer-reviewed yet. My take: Autor calls AI "a performance equalizer, but a skill-disequalizer." Seniors used it to audit logic they already understood. Juniors got to skip the blank page, which is exactly where judgment gets built. The one thing I can't find in the paper: what happens to juniors after year two. If you manage juniors: do you ban AI on first drafts, or let them use it and test them without it?