Chris Kennedy

4.8K posts

Chris Kennedy banner
Chris Kennedy

Chris Kennedy

@c3K

{Psychiatric, causal, psychometric, robotic, neuro} AI. @MGHPrecisionPsy @MGHPsychiatry @HarvardMed; @UCBerkeley biostat PhD

Boston, MA Katılım Ağustos 2007
4.1K Takip Edilen1.4K Takipçiler
Sabitlenmiş Tweet
Chris Kennedy
Chris Kennedy@c3K·
Latest suicide pub (w/ @drkatebentley @TaylorABurkePhD @jorsmo @DrJBStephens) provides strong evidence to 1) counter the common claim that clinicians can't predict suicide risk, demonstrating the value of clinical judgment, yet also 2) massive improvement through ML on SRA items
JAMA Psychiatry@JAMAPsych

While clinicians can stratify suicide risk above chance levels, predictive accuracy for future suicide attempts significantly improves when using machine learning to incorporate comprehensive clinical assessment data. ja.ma/3G3Niy2 @drkatebentley

English
0
7
16
3.2K
Jason Steen
Jason Steen@Steen_Dr·
@JohnHolbein1 PNAS has two submission methods. member contributions (2 papers annually within your expertise) or full outsider peer review. doesnt mean its not good research, but there are two distinct types of paper in PNAS.
English
1
0
3
9K
Chris Kennedy
Chris Kennedy@c3K·
@johnmyleswhite For guaranteed results, hooks are necessary, eg via cc-sessions - memory is def not the way to prevent actions
English
1
0
1
64
John Myles White
John Myles White@johnmyleswhite·
Am I the only person who finds Claude Code's memory system to be too unreliable to use? I've learned to encode everything as skills and deny permissions instead because CC having a memory saying "don't ever do X" basically just gives me a 50/50 probability that it won't do X.
English
8
1
12
2K
Chris Kennedy
Chris Kennedy@c3K·
@GlobePulses @newstart_2024 We’re at about 1 hr/week, only on weekends, and daniel tiger or nature/science documentaries primarily. 0 most weeks, but can be a little more (eg 4/week) if we’re extra exhausted for some reason. Parent should interact with child during this so it’s not entirely passive.
English
0
0
1
201
GlobePulses
GlobePulses@GlobePulses·
The scary part isn’t screens are evil, it’s that MRI studies are already showing white matter changes in 3–5-year-olds at around 2 hours a day in the same tracts tied to language and early literacy. If you’re a parent, what’s your actual screen rule in your house right now (and why)?
English
11
1
15
12.5K
Camus
Camus@newstart_2024·
Shocking new MRI study on 3–5 year olds: Just 2 hours of interactive screen time per day is linked to measurable loss of white matter in the brain. Professor Mike Nagel (University of the Sunshine Coast): “White matter is myelin — it insulates axons like plastic on a wire. Deficits in myelin early in life mean deficits in neural connectivity.” The more screen time, the greater the white matter loss — especially in areas tied to language development and literacy. Nagel’s first reaction (as a researcher and father): “Wow… I was not anticipating seeing anything like that. It hadn’t occurred to me that something as little as two hours a day was having such a profound effect.” Parents: Has this changed how you think about screen time limits for young kids — or do you think the risks are overstated?
English
279
3.1K
9K
2M
Graham Neubig
Graham Neubig@gneubig·
This is definitely something to be aware of both for benchmark builders and users IMO. For longer-running, more difficult tasks, the differences between which agent you use can be big, like a 10% gain in success rate when going from Claude Code to OpenHands.
Wayne Chi@iamwaynechi

@OpenHandsDev OpenHands leads to such a sharp improvement on Codex 5.1 (33.3% ➡️ 43.2%) and Claude Sonnet 4.5 (34.1% ➡️ 45.5%) that it makes me wonder if some of the other leaderboards would see improvement from trying out different frameworks as well.

English
2
0
42
5.9K
Karthik Bala
Karthik Bala@Karthik87403069·
@juliarturc @hopes_revenge Consciousness may as well be binary, which is what panpsychism claims. It may sound woo-woo at first, but even modestly rigorous reasoning and inner exploration through meditation lead to a noteworthy likelihood of there being an on/off switch.
English
1
0
0
45
hope hopes hoping
hope hopes hoping@hopes_revenge·
1) why is there something rather than nothing ? 2) how did life emergence out of non-life? 3) what is consciousness? thank you
English
205
23
532
30K
Silvi Rouskin
Silvi Rouskin@silvirouskin·
@c3K It’s literally called “a trainee” 💀
English
1
0
1
308
Silvi Rouskin
Silvi Rouskin@silvirouskin·
Academia math: train till 40, finally snag a job, and folks want you retired by 65. So that’s 40 years of training for 25 years of mildly paid chaos. Absolutely not. I’m staying in this job at least as long as I trained—or until I drop dead mid–faculty meeting, whichever comes first.
Silvi Rouskin tweet media
English
24
12
179
18.9K
Chris Kennedy
Chris Kennedy@c3K·
@j_g_allen May need to revise your mental model. When an AI scans files it generally isn’t training on them (requires significant GPU resources) nor remembering them (eats up context window). There is still a risk of sharing/using specific files when they are being read, or prompt injection
English
1
0
1
80
Joseph Allen
Joseph Allen@j_g_allen·
the leaking of secret corporate IP is coming hard for businesses... Simple Scenario -employee is using AI to work on a work project, but can't find a file or wants to synthesize some historical meeting notes -employees uses AI to scan all the company files to find a doc or ppt or something benign -the AI then reads *ALL OF THE COMPANY FILES*, looking for this one little thing, but now it knows EVERYTHING -the employee connects their chatbot to moltbook
Joseph Allen@j_g_allen

Well, the moltbook thing certainly changed my thinking on one thing... I was using antigravity to building a web app for work. Very straightforward project. But it hit a snag, so, to problem solve, the AI asked to *start searching* ALL of my files (yeah, like kids stuff, medical, business, financials, tax docs, PASSWORDS, credit card info...). I think prior to moltbook I was thinking, "well, yeah, that'd be useful to have this AI thing know where all my work stuff is so I can be more efficient..." But then with moltbook, you realize your AI agent, trained on your files, emails, and queries, could very easily be out in the real world...

English
3
0
11
2.1K
Chris Kennedy
Chris Kennedy@c3K·
@stanislavfort @jeremyphoward Plus, robots exist, right? So why exactly can’t they run experiments? A lot of experiments are virtual anyway, or can be run using online tools.
English
0
0
0
6
Stanislav Fort
Stanislav Fort@stanislavfort·
I don't think it's the case that "LLMs can't run experiments". a) with a bunch of tools they can (eg LLM connected to a remote telescope), b) they definitely can run simulations which can be as good sometimes (eg in cosmology) & c) they can learn fast simple models of the world & run virtual experiments on them in their mind (eg AlphaFold)
English
2
0
10
794
Jeremy Howard
Jeremy Howard@jeremyphoward·
For those that hope (or worry) that LLMs will do breakthrough scientific research, I've got good (or bad) news: LLMs are particularly, exceedingly, marvellously ill-suited to this task. (if you're a researcher, you'll have noticed this already) Here's why🧵
English
114
580
4K
1M
Pooja Algikar
Pooja Algikar@algikarpooja·
Physicists: because it feels like they expand what’s even possible. I’m no physicist, so I might be missing a lot (by a huge margin). Mathematicians: this goes hand-in-hand with physics so much. Like @miniapeur said, there isn’t a single clean traced line that divides the two. Spies: the tolerance capacity… couldn’t imagine. I visited a spy museum once and it’s been an all-cry museum for me since then. Cognitive scientists: AGI bringers. If anyone can do it, it’s them. (could be way off.) Economists: mind-boggling causal chains. I can see how anyone could go mad. Philosophers: don’t have to make the case for them. Statisticians: living in the space of uncertainty, coming up with approaches using probability distributions… again, mind-occupying.
English
1
0
1
91
Pooja Algikar
Pooja Algikar@algikarpooja·
First seven professions you have dying respect for? My list: 1. physicists 2. mathematicians 3. cognitive scientists 4. economists 5. philosophers 6. spies 7. statisticians
English
4
1
9
2.1K
Chris Kennedy retweetledi
Rex "garbage in" Douglass Ph.D.
Rex "garbage in" Douglass Ph.D.@RexDouglass·
WorstDaySoFar [👇🏻link] Daily summary and tracking of US mass rights violations and democratic collapse Disappearances, warrantless detentions, dragnets, political prosecutions, and 250+ other types of events Fully automated and community funded
English
1
12
24
19.7K
Chris Kennedy retweetledi
Rex "garbage in" Douglass Ph.D.
Rex "garbage in" Douglass Ph.D.@RexDouglass·
Worst Day So Far SITREP - Authoritarian Consolidation Last Update Jan 24, 7:11 PM · LLM Dec 1, 2025–Jan 23, 2026 (54d searched) New Today Tactics: Scene access denial -- federal teams block state investigators after lethal force incidents (Minneapolis) Violence/Detention: Border Patrol lethal force -- urban shooting escalates risk during routine immigration operations (Minneapolis) • Stun grenades and tear gas -- crowd control spikes after shootings, raising bystander injury risk (Minneapolis) New Yesterday Tactics: Generic POLICE vests -- misidentification enables federal teams to evade accountability during grabs (Minnesota) Lawfare: Grant drawdown toggles -- federal health funding access paused to enforce priority compliance (US) • Sealing special-counsel report -- privilege claims used to suppress scrutiny of executive misconduct (US) Disinformation: National parks passes with leader portrait -- voiding altered passes coerces visible loyalty (US) • Geofenced ICE recruitment memes -- wartime branding accelerates staffing for enforcement surges (US)
Rex "garbage in" Douglass Ph.D. tweet media
English
0
1
2
435
Dylan Armbruster
Dylan Armbruster@dylanarmbruste3·
@AnilMakam If you aren't talking about a specific study, then I take this to be an empty meta take that shouldn't be taken seriously. If you want to cry confounding, tell us the confounders that would need to be controlled for that have a larger effect than the effect of treatment.
English
1
0
1
29
JFPuget 🇺🇦🇨🇦🇬🇱
JFPuget 🇺🇦🇨🇦🇬🇱@JFPuget·
A difference between codex cli and claude code is that the latter has a jupyter notebook tool. I thought this might be a weakness of codex and looked for existing skills. Found one, tried it. Well, adding that skill to codex makes it work worse with notebooks: takes longer, and sometimes outputs corrupted cells. Are people trying the skills they produce? If so, how do they do it? I don't want to flame that particular skill, maybe I was unlucky. My point is more on how to evaluate skills. Anyway, here it is for those willing to deep dive: github.com/narang99/jupyt…
English
8
0
22
3K
Silvi Rouskin
Silvi Rouskin@silvirouskin·
That’s actually very helpful. I agree her refusal was super odd. I wrote to her again, but she replies once in two moths ! I’m assuming she did not what to specify the exact brand of supplements but it shouldn’t matter it’s the brand I’ve been using to get my lab results to normal!
English
1
0
1
130
Silvi Rouskin
Silvi Rouskin@silvirouskin·
I am upset … I have diagnosed iron deficiency anemia, and Hashimoto, and inability to absorb B12 and diagnosed panic disorder and I take all these vitamins and supplements to sustain myself (through various vendors) and my PC doctor refused to write a letter so I can use it for my HSA. 💀 like just plain said NO. Is that even allowed 😆 she has to agree to write a letter at least describing my diagnosis, or do I just show a copy of my lab results ? What a world, clearly noone cares about my health ….
English
14
1
81
16.4K
Chris Kennedy
Chris Kennedy@c3K·
@silvirouskin @Alex_Zapata001 If you use One Medical it's like a week - that's what I went with when MGH told me it'd be like 9 months to see a PCP. And other than the annual check-up you can mostly do virtual visits.
English
1
0
1
110
Chris Kennedy retweetledi
Ben Van Calster
Ben Van Calster@BenVanCalster·
Our overview of and guidance for performance measures to evaluate medical AI is finally out! - Stop bashing AUROC - Calibration + clinical utility are key - Plot risk distributions - Classification measures are improper thelancet.com/journals/landi…
English
6
35
113
19.5K

Keşfet

@Steen_Dr @JohnHolbein1 @johnmyleswhite @GlobePulses @newstart_2024 @gneubig @Karthik87403069 @juliarturc