Sebastian J Zhao

7 followers · 11 following · 1039 views

on the atlas — 50

highlights — 243

Unlike ales, lagers tend to be crisp and dry
Beer 101: Types and styles - The Beer Store
The type of yeast is what makes ales and lagers different,
Beer 101: Types and styles - The Beer Store
designed for or understood by the specially initiated alone
ESOTERIC Definition & Meaning - Merriam-Webster
nder this metric, spawning more subtasks only helps if it shortens the critical path.
Kimi K2.5 Tech Blog: Visual Agentic Intelligence
uantization noise remains static throughout the process, lacking the adaptability needed to enhance exploration at critical phases
[2510.11696] QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
promot- ing broader exploration of potential actions or tokens in RL by increasing entropy
[2510.11696] QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
To avoid this drawback, our analysis surprisingly reveals that quantization noise, with precise control, can benefit RL by increasing policy entropy (Fig.3).
[2510.11696] QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs
LoRA performs equivalently to FullFT for reinforcement learning even with small ranks. We find that RL requires very low capacity, a result we anticipated based on information-theoretical arguments.
LoRA Without Regret - Thinking Machines Lab
For supervised fine-tuning on small-to-medium-sized instruction-tuning and reasoning datasets, LoRA performs the same as full fine-tuning.
LoRA Without Regret - Thinking Machines Lab
Since LoRA trains an adapter (i.e., the A and B matrices) while keeping the original weights unchanged, a single inference server can keep many adapters (different model versions) in memory and sample from them simultaneously in a batched way.Punica: Multi-Tenant LoRA Serving (Chen, Ye, et al, 2023) Modern inference engines such as vLLM and SGLang implement this feature.
LoRA Without Regret - Thinking Machines Lab
On the flipside, alcohol usage is also linked to QLC.
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
Words relating to time (“night”; “weekend”; “morning”; “early”; “day”) and work (“work”; “working”) had the highest frequency and correlation strengths.
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
The terms associated with the QLC based on the holistic model of early adult crisis are: Stuck; Trying; Leave; Change; Unemployed; Lonely; Hopeless; Overwhelmed; Unfair; Fail; Coping; Failing; Debt; Meaning; Trapped; Try; New; Identity; Sacked; Money.
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
Social media postings that relate to actual life events and experiences can be argued to serve a developmental function, which is to represent and reify the passing of time into a simplified and publicly documented life story that can help the individual create a meaningful ongoing narrative of how their life is changing (Rettberg, 2009).
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
locked out of adult commitments (being unable to find work or love), or the feeling of being locked in to life roles that are then experienced as a poor fit for one’s identity,
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
Early adult crisis episodes typically occur toward the latter end of the life stage of emerging adulthood, and last approximately a year
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
Epidemiological data shows that most young adults in Western countries now choose to wait for a decade or more after turning 18 before having children, or before starting a marriage or civil partnershi
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
Users who refer to a QLC were found to post more about feeling mixed emotions, feeling stuck, wanting change, career, illness, school, and family.
Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter - PMC
In losing Hae-mi and murdering Ben, Jong-su undoubtedly suffers, but in many respects, he is the author of his own pain.
Watching the World Burn | Burning (2018) - Bright Wall/Dark Room
Yes, Hae-mi did not disappear or die by his hand, but considering Jong-su’s many reflexive cruelties throughout the film, she died numerous little deaths unbeknownst to him.
Watching the World Burn | Burning (2018) - Bright Wall/Dark Room
you construct an angry jeremiad out of your nostalgia for the props of the old reality and the architecture that's been demolished.
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
gifts" of imagination and transcendence, but enhance them
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
The borders separating simulations from reality are important because they remind us of the limits that make dreams of technological transcendence dangerous fantasies.
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
Even within the boundaries of simulations, material intractability often breaks in.
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
Only when these boundaries do not exist, or cease to signify that one has left the simulation and entered reality, does the dreamscape that Baudrillard evokes shimmer into existence.
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
The Iowa farmer who has spent the day inspecting his seed corn, feeding his hogs, and spreading manure on his garden will not be easily persuaded that he lives in a world where it is no longer possible to distinguish between simulation and reality.
In Response To Jean Baudrillard (Hayles, Porush, Landon, Sobchack, Ballard)
They speak of the pursuit of “social justice” when what they really mean is abolishing free markets. They speak of “equity” when what they really mean is illegal and highly unpopular racial quotas.
What Does Postmodernism Really Amount To? - Claremont Review of Books
We’ve seen now in soaring crime rates the danger of indiscriminate hostility to all existing customs and dogmas.
What Does Postmodernism Really Amount To? - Claremont Review of Books
it’s hard to find any significant original thought in postmodernism.
What Does Postmodernism Really Amount To? - Claremont Review of Books
Derrida and his followers completely misunderstood what Saussure was saying here.
What Does Postmodernism Really Amount To? - Claremont Review of Books
for instance, maybe em-dashes just read more conversational, so they were preferred by RLHF-ers
Why do AI models use so many em-dashes?
if em-dashes are common because they’re a feature of late-1800s/early-1900s writing, why doesn’t AI prose read more like Moby-Dick?
Why do AI models use so many em-dashes?
State-of-the-art models rely on late-1800s and early-1900s print books for high-quality training data, and those books use ~30% more em-dashes than contemporary English prose.
Why do AI models use so many em-dashes?
multi-GPU inference sig- nificantly increases the hardware cost of deploying MoEs.
MiLo: Efficient Quantized MoE Inference with Mixture of Low-Rank Compensators
what they actually mean is that they cannot imagine slumming their twenties on less than six figures a year!
it's time to start bullying young Palantir employees
But it’s worth noting that few low-income students statistically go into fields like consulting.
it's time to start bullying young Palantir employees
Your ancestors’ wildest dreams was not to work at lockheed martin stop lying on them folks.
it's time to start bullying young Palantir employees
But it’s also just a feature of being a mortal on earth.
it's time to start bullying young Palantir employees
fear of post-grad risk and uncertainty — incidentally, the very conditions in which bad bitches are made
it's time to start bullying young Palantir employees
Her co-founder is a world-famous climate activist who once publicly partook in a Black Friday strike with Jane Fonda against consumerism in 2019. Again, you cannot make this up!)
it's time to start bullying young Palantir employees
neoliberalism, the bitch-ass belief
it's time to start bullying young Palantir employees
“Fifteen years earlier, graduates of the country’s most elite colleges had often been concerned with trying to improve the state of the world. Now, the focus was different: How can I be as financially successful as possible?”
it's time to start bullying young Palantir employees
"We think what happened is that the cats sort of domesticated themselves,"
A Brief History of House Cats
All domestic cats, the authors declared, descended from a Middle Eastern wildcat, Felis sylvestris, which literally means "cat of the woods." Cats were first domesticated in the Near East, and some of the study authors speculate that the process began up to 12,000 years ago.
A Brief History of House Cats
During the generation process, as the context continually evolves and may include mispredicted tokens, the confidence of otherwise stable tokens can fluctuate or even regres
CreditDecoding: Accelerating Parallel Decoding in Diffusion Large Language Models with Trace Credits
For each sequence in the batch, if its minimum score within the current window min( S W j ) falls below a predefined threshold τ thresh , a refinement cycle is initiated for its window W j
Review, Remask, Refine (R3): Process-Guided Block Diffusion for Text Generation
ather than exploring uniformly, we target medium-confidence positions near an information level c info ≈ 0 . 2 .
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models
hen a previous block lacks high-confidence tokens (none exceed threshold C ), ETE unmasks the highest-confidence masked token in that block
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models
ETE allows high-confidence tokens in earlier blocks to be unmasked in parallel with the current block
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models
e assign a uniform sampling budget of N diffusion iterations per block and unlock the next block once this budget is exhausted. This prevents the algorithm from stalling in low-throughput regimes where blocks are nearly complete and few tokens remain to be decoded
From Bits to Rounds: Parallel Decoding with Exploration for Diffusion Language Models