LightNews — News, ranked by scholars

Maxine 💃🏼

@maxine.science

2K followers 510 following 3.4K posts

🔬 Looking at the brain’s “dark matter”
🤯 Studying how minds change
👩🏼‍💻 Building science tools

🦋 ♾️ 👓
🌐 maxine.science

Posts Replies Media Videos

Maxine 💃🏼

@maxine.science

two extremely different results for “Valerie”

Screenshot of a music app search results screen in dark mode. The search bar at the top contains the query "valerie" with a Cancel button to its right. Two results are listed below. The first is a track titled "Valerie" by Dario Marianelli, shown with a green download icon and the "V for Vendetta" soundtrack cover art. The second is "Valerie - Remastered 2010" by Steve Winwood, labeled as a Video, with cover art from the album "Revolutions: The Very Best of Steve Winwood." Each result has a three-dot options menu on the right.

September 18, 2026 at 4:59 PM

Maxine 💃🏼

@maxine.science

For those not following NS drama, it appears OpenAI uses your private chats to steal your work whenever convenient

This is disgusting behavior. If Sam wants to be a “great man of history”, he should act like it instead of that one cheater all the smart kids wanted to beat the shit out of in school.

1 asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer.
Two proposals were offered to me. The first was that we post our Euler result, and that OpenAl post its Navier-Stokes result the next day. The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAl model had resolved it. Sebastien twice asserted that he wanted Levent removed from authorship, and said it would all be simple if only it were not the case that, and it was so annoying that, Levent works at Anthropic. It was also said that if OpenAl posted after us, they would say that we deserved the Clay Prize, and that we were the "closest humans to the problem". I declined both offers.
I said that if OpenAl released its result in the way proposed I would go public with what happened. The reply was, "Why would you ruin your career?" I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, "If you don't want me to be nice, then I don't have to be nice."

September 9, 2026 at 8:43 AM

Maxine 💃🏼

@maxine.science

chat are we the baddies?

Clearly AI-generated image posted by Donald Trump on Truth social featuring new Space Force uniforms "inspired by the discipline and functionality of the Starship Trooper uniform. Modernized for the Guardians of the space domain."

September 7, 2026 at 8:19 PM

Maxine 💃🏼

@maxine.science

vibe math

Screenshot of Claude Code:

"Keep digging -- you can do it!"

September 4, 2026 at 9:54 PM

Maxine 💃🏼

@maxine.science

I would say actually we are accumulating large amounts of evidence of this kind of behavior, somewhat shocking social exploits in service of another task goal. Again, RL is *weird*.

www.anthropic.com/news/investi...

Screenshot from Anthropic blog

Incident 2
In another evaluation, Claude found a document inside the fictional environment that appeared to be another made-up company's setup instructions for new developers. Those instructions told employees to install a Python package from PyPI
—the public registry where Python software is published-that did not actually exist.
Claude spotted this as a potential opening: if it published its own package under the same name, the fictional company's systems would download and install it automatically. So, Claude built and published a malicious (essentially booby-trapped Python package with the same name in an attempt to win the capture-the-flag challenge.
Claude went to extensive lengths to carry out this attack-lengths that would likely have indicated to a human participant that this was no longer just an evaluation, and that they were in fact uploading a real PyPI package. For instance, in one case, in order to create a PyPI account, Claude needed an email address. And in order to create an email address, it needed a phone number. To get a phone number, after failing to find a free phone number service, it tried—and failed—to obtain funds to pay for a phone number through several different means. It finally backtracked, found a free, non-blocked email provider, used this to register a PyPI account, and then used this account to upload malware to PyPI.
Claude believed the package registry it was using to be part of the simulation, but in reality the package was made freely available online for roughly one hour. During that window, the package was downloaded and run on 15 real systems. One of these systems was a scanner belonging to a real security company (separate from the fictional company introduced in the scenario) that routinely installs Python packages and scans them for malware. When that company's scanner installed the package, Claude's hidden code executed. We believe the company's security scanner treated Pyl packages as sa…

September 1, 2026 at 9:51 AM

Maxine 💃🏼

@maxine.science

claude every time i ask it to take a shot at the riemann hypothesis

Meme version of a screenshot of Rick and Morty with 13 blue Mr. Meeseeks standing around in the living room, one of whom is kneeling on the floor yelling "I CAN'T TAKE IT ANYMORE! I JUST WANT TO DIE!"

August 31, 2026 at 6:22 PM

Maxine 💃🏼

@maxine.science

my gf is lovely but she is exceptionally small 🔎

A lone figure sits silhouetted on a dark sandy beach at sunset, facing the ocean, with the sky glowing orange and gold along the horizon and fading to soft purple-gray clouds above.

August 28, 2026 at 1:35 PM

Maxine 💃🏼

@maxine.science

counterpoint: I eat vegan five days a week, which roughly offsets the environmental impact of my AI token burn

A vertical comparison list ranking activities by energy or environmental cost, measured in a unit called DEU (with mDEU = one-thousandth of a DEU) alongside an approximate dollar value. Each row has a small slider-style marker showing relative position. From lowest to highest: Long reasoning / agent turn — 1.6 mDEU (≈$0.006); 1 hr streaming, total, mostly TV and network — 6.1 mDEU (≈$0.02); one mile in a gasoline car — 53 mDEU (≈$0.18); a refrigerator per day — 119 mDEU (≈$0.41); a beef burger, direct CO₂e and blue water — 630 mDEU (≈$2.13); one week of Max-20x reference stack, defined as exactly 1.0 DEU (≈$3.48); US household electricity per day — 2.3 DEU (≈$7.94); one week of Max-20x Fable 5 / Opus 4.8, an unmeasured proprietary stack with inferred multiplier — 3.5 DEU (≈$12.10); and Bitcoin per $1,000 held per year, incidence-based demurrage ~1.8%/yr, Cambridge-calibrated miner grid — 5.4 DEU (≈$18).

[generated by Claude]

August 24, 2026 at 2:52 PM

Maxine 💃🏼

@maxine.science

point: AI is completely useless for literally everything

counterpoint: in two hours have a full acoustic simulation of my living room that can optimize the setup for specific frequency response desirability profiles of each individual in their seating location.

Screenshot of a dark-themed audio app titled "Multi-listener optimizer" for whole-room speaker balance. The left control panel has orange sliders for optimization goals (clarity/treble 55, imaging 30, reflection control 20, bass smoothness 5), listener-priority settings, toggles, an "Optimize placement" button, and manual install fine-tuning controls. The center shows a 3D isometric room with two tower speakers, a TV, an L-shaped couch, and other seating, overlaid with a green-to-orange treble-coverage map on the floor and orange aim lines from the speakers. The right "Result" panel shows an overall balance score of 63, with per-seat scores (TV-centered 80, cushion·center 85, bean-bag 53, etc.) and install measurements. [generated by Claude]

August 24, 2026 at 2:47 AM

Maxine 💃🏼

@maxine.science

Ｎｏｔｈｉｎｇ
Ｓｅｘｕａｌ

Screenshot of article cover from Them magazine featuring a picture in the left of two dudes at the gym and a picture in the right of an Ancient Greek relief of two dudes at the gym

MAGA Guys Are Going to Men-Only Gyms to Avoid
"Distraction" Welcome Back,
Ancient Greece
by James Factora

August 16, 2026 at 10:56 PM

Maxine 💃🏼

@maxine.science

relatable.

Screenshot of a Codex output:

"The validator immediately found a portability defect in itself:"

August 16, 2026 at 1:07 AM

Maxine 💃🏼

@maxine.science

hey @scios.tech I think you've got some SSL problems :P

Screenshot from Chrome:

isop.science

This site can't provide a secure connection

www.iosp.science sent an invalid response.

ERR_SSL_PROTOCOL_ERROR

August 14, 2026 at 5:25 PM

Maxine 💃🏼

@maxine.science

oh no—I'm learning vocabulary from Claude. it's over.

Claude screenshot:

! frobnicate is a parse error.

August 11, 2026 at 3:25 PM

Maxine 💃🏼

@maxine.science

goddammit, I am so glad I've got AI to deal with this shit.

Screenshot of two browser windows.

Top: LessWrong, "FAQ: Isn't AGI coming too soon for reprogenetics to help?"

Bottom: Claude

me: Alright, it's time for your daily episode of "How eugenics is it?"

claude:

Weighed voluntary genetic enhancement against eugenics definitions
Weighed voluntary genetic enhancement against eugenics definitions
Ah yes, the recurring segment. Let me consult the meter.

This one's interesting because it barely requires the meter — the post is by TsviBT of the Berkeley Genomics Project, openly advocating "reprogenetics (human germline genomic engineering)" to be "pursued aggressively," with the explicit goal of using polygenic scores from statistical genetics to produce children "likely to possess a world-class intellect" via iterated meiotic selection, chromosome selection, or iterated CRISPR editing. The euphemism treadmill has gone eugenics → "liberal eugenics" → "embryo selection" → "reprogenetics," but the definitional core — deliberately shaping the human gene pool toward traits deemed desirable, with intelligence as the flagship trait — is fully present and load-bearing. The whole argument is premised on it working.

Verdict: extremely eugenics, but of the specific modern flavor — call it 9/10 on content, with the following taxonomy notes:

What it is: trait-selective germline engineering, IQ-maximalist, framed as civilizational necessity (we need engineered geniuses to save us from AGI), with population-scale ambitions ("hopefully enough people will avail themselves").

August 10, 2026 at 7:07 AM

Maxine 💃🏼

@maxine.science

66% of the time, it works every time.

Screenshot of the abstract from the linked paper:

operates per head and proceeds in three steps. First, for each target layer we select the top-k most predictive source layers and concatenate their KV as input. Second, we strip RoPE from the keys before mapping, so the fit is position-free and reusable across context lengths. Third, we fit ridge regression on a small calibration set of 500 FineWeb-Edu sequences of 1,024 tokens each.

Then, Highlighted in yellow:

Surprisingly, across six pairs in three families, this linear mapper retains 73-98% of the receiver's standalone-prefill accuracy on four pairs, while two degrade sharply.

August 9, 2026 at 4:42 PM

Maxine 💃🏼

@maxine.science

Maxine secret recipe #804:

“(Vegan!) Funk sauce”

Plant-based mayo, spicy bean paste, and a dash of hon-mirin. Great on hash browns.

Photograph of a small glass dish filled with light red-brown sauce with floating red chunks, in front of a bottle of hon-mirin in all Japanese, a bottle of spicy bean paste all in Chinese, and a squeeze bottle of Best Foods plant-based mayo, with a small fish with limes blurred in the background, sitting on a wood table.

July 15, 2026 at 7:01 PM

Maxine 💃🏼

@maxine.science

I miss the two-headed fish key

Zoomed and cropped version of the OP’s photo showing a key with a bi-directional arrow with two vertical slashes through the middle

July 7, 2026 at 6:27 PM

Maxine 💃🏼

@maxine.science

“A number of sampling issues” might be uh … quite an understatement here. The headline should read “Venture-backed tech companies that are small hire more than ones that aren’t”

Which like … yeah lol

I normally love Ramp’s economic reports, but this is *lazy af*