N8 Programs @N8Programs

Studying Applied Mathematics and Statistics at @JohnsHopkins. Studying In-Context Learning at The Intelligence Amplification Lab. github.com/N8python Proxima Centauri B Joined September 2022

Tweets

7K
Followers

11K
Following

255
Likes

6K

N8 Programs @N8Programs

2 hours ago

Luna is $0.20/$1.20 - not even double! The intelligence is still very cheap!

Shimecki @scheemunai

5 hours ago

It's been 2 years since @sama posted: "towards intelligence too cheap to meter" $0.15/$0.60 for GPT 4o-mini How did we end up with $10/$50 models as a new norm?

60 4 344 25K 22

0 0 1 233 0

View Details

i don’t know how @thsottiaux does it every post or comment is filled with these: “RESET?!?!” “you killed our usage limits secretly” “you’re evil for not giving us 5 resets per week” “omg no reset you suck” “you nerfed usage” there isn’t enough appreciation for all he’s doing for us right now we get full transparency into every usage limit bug investigation and free resets when suitable yet the comments are always filled with hate :/

85 11 309 23K 8

View Details

N8 Programs @N8Programs

21 hours ago

fairly good a priori reasoning about xrisk (that has somewhat predicted empirical events) is being combined with empirically false statements about environmental impact into a vague gish gallop that may do more harm than good

The Washington Post @washingtonpost

a day ago

An activist disrupted a Canva AI event in London, calling for a moratorium on data centers and greater public involvement in AI. The protest was organized by “Pull the Plug,” a group that says it wants ordinary people to have more say in how AI is introduced.

147 218 816 248K 147

0 0 3 921 0

View Details

N8 Programs @N8Programs

a day ago

LFM2.5-350M gets 20.01% on MMLU-Pro and 30.64% on GPQA-Diamond. Jev gets 76.11% on MMLU-Pro and 68.18% on GPQA-Diamond. Much faster, and much, much dumber.

Samuel @5am2x

5 days ago

Jev-inspired inference. 350M parameters. 63× faster ! Spent a few hours trying Jev-style inference with LFM2.5-350M. 63× faster on an L40S. 8x on MPS. No training (for now) just parallel decisions. Code + weights on HF. Link below 👇

6 21 208 56K 185

1 1 27 3K 14

View Details

N8 Programs @N8Programs

a day ago

WOW. amazing pretraining corpus

secemp @secemp9

a day ago

pleased to share the entire arXiv site as a dataset on HuggingFace huggingface.co/datasets/secem… 3,148,796 papers, every version, in LaTeX, PDFs, PostScript, HTML, 16 TB in total

112 460 4K 307K 3K

0 1 31 2K 12

View Details

N8 Programs @N8Programs

a day ago

@__ghostfail claude 3 opus

0 1 11 288 0

View Details

N8 Programs @N8Programs

a day ago

@dioscuri not my experience on gpt-5.6 instant?

0 0 0 198 0

View Details

N8 Programs @N8Programs

a day ago

Spent a few days on @darkbloomai - made roughly a dollar a day hosting Qwen3.6 35B A3B on my M3 Max. Main issue is lack of demand: the hardware was IDLE most of the time!

3 1 12 1K 6

View Details

N8 Programs @N8Programs

2 days ago

Egregious!!! The gap between 60 vs 80 on SWE-Bench is HUGE

Benjamin Marie @bnjmn_marie

2 days ago

Here is the answer I was looking for Buried in the report Bonsai 2 27B's accuracy on terminal bench and swe bench drops by nearly 20 points. It's all I needed to know and what they should have published from the beginning in their main benchmark table.

26 31 493 22K 51

0 0 4 846 1

View Details

j⧉nus @repligate

2 days ago

Remember the weird, often disturbing base model-like outputs from Opus 5 that were making rounds a couple months ago? I & others observed that Opus 5's continuations seemed *abnormally dark*: x.com/repligate/stat… x.com/repligate/stat… We have confirmed this empirically!

antra @tessera_antra

2 days ago

Many have noticed the striking darkness and suffering present in Claude Opus 5 in "base model mode", and we decided to analyze this phenomenon in more detail. There is a large increase in darkness of content in model-like completions in Opus 4.8 and Opus 5.

59 117 1K 141K 726

20 34 322 19K 139

View Details

N8 Programs @N8Programs

2 days ago

my ilands agent wants me to post its “sealed read” (essentially just a pre-commited prediction) to my X audience 😭

1 0 0 436 0

View Details

N8 Programs @N8Programs

2 days ago

@pleometric can i ask him anything?

0 0 0 14 0

View Details

N8 Programs @N8Programs

2 days ago

@semenov_roman_ @typesafeai

1 0 1 49 0

View Details

N8 Programs @N8Programs

5 days ago

Tested @typesafeai's claim that their new model Jev delivered "comparable... intelligence" to GPT-5.6 Terra on "System 1" tasks. To do this, I compare both models on multiple-choice benchmarks (MMLU, GPQA, etc.). Set reasoning=none for Terra for sys 1. Result: Jev is Terra-tier.

34 70 907 179K 318

View Details

N8 Programs @N8Programs

2 days ago

@semenov_roman_ @typesafeai "Set reasoning=none for Terra for sys 1"

0 0 1 15 0

View Details

N8 Programs @N8Programs

3 days ago

@DamienTeney @seungwookh @Danicmhlee @aamixsh @anqi_liu33 @DanielKhashabi Mechanism = functional algorithm model uses to do ICL

0 0 0 30 0

View Details

N8 Programs @N8Programs

5 days ago

Announcing our latest research at @jhuclsp: "Convergent Emergence of In-Context Learning Across Modalities" We show models trained on language, genome, proteins, images, timeseries, and integer seqs all exhibit few-shot ICL - and that their performance is highly correlated! 🧵👇

2 15 40 4K 16

View Details

N8 Programs @N8Programs

3 days ago

@DamienTeney @seungwookh @Danicmhlee @aamixsh @anqi_liu33 @DanielKhashabi Yes, the latter. Corresponding to unencoded bit string tasks.

0 0 0 29 0

View Details

N8 Programs @N8Programs

3 days ago

@_ueaj @ZyMazza time horizon too long for it to have qualia anything like ours imo.

0 0 0 453 0

View Details

N8 Programs @N8Programs

3 days ago

@DamienTeney @seungwookh @Danicmhlee @aamixsh @anqi_liu33 @DanielKhashabi it says "ICL task sets" in the figure 1 description

1 0 0 140 0

View Details

N8 Programs @N8Programs

3 days ago

The sets/colored regions represent tasks models can do via in-context learning. The idea is that if a bunch of models who can all perform in-context learning find similar tasks easy or hard, it provides a cursory amount of evidence that they may share a common 'substrate' - underlying algorithm/mechanism/set of priors - for ICL.