bhadass
they're almost certainly going to replace all the board memebers with political loyalists. the board members served six year terms specifically so they'd span multiple administrations and stay independent.
firing them all at once lets you stack the entire board with people. it's not about making science better, it's about removing the people who'd say no.
the "benefit" from his perspective is the same playbook trump admin has been running across every federal agency, he wants to replace independent experts with loyalists, remove checks on executive power, and redirect spending toward admin priorities.
the board members served six year terms specifically to insulate science funding from political cycles. that's a feature to everyone else and a bug to this administration.
could an AI native competitor eventually eat their lunch? sure. but "no new sane company wants to use their products" is a stretch when their customer count is literally still growing double digits.
i would not say "writing on the wall" at all.
when someone invents a new scheduling algorithm or a new concurrent data structure, it's usually based on hunches and empirical results (benchmarks) too. nobody sits down and mathematically proves their new linux scheduler is optimal before shipping it. they test it against representative workloads and see if there is uplift.
we understand transformer architectures at the same theoretical level we understand most complex systems. we know the principles, we have solid intuitions about why certain things work, but the emergent behavior of any sufficiently complex system isn't fully predictable from first principles.
that's true of operating systems, distributed databases, and most software above a certain complexity threshold.
i'm guessing most of the gains we've seen recently are post training rather than pretraining.
funny thing is quaternions had that exact same energy in the computer graphics community for years. after ken shoemake introduced them to CG in 1985, there was a long period of "why are we using euler angles like cavemen when this exists??". now quaternions are well known tooling for people in graphics and the mystique has worn off at least in that community.
Lots of stuff you could do. Adjust the system prompt, add guardrails/filters (catching mistakes and then asking the LLM loop again), improve the RAG (assuming they have one), fine tune (if necessary), etc.
ZIRP (especially the "double tap" ZIRP in 2021/2022) created this monster (bootcamp devs getting hired, big tech devs making "day in the life of" tiktok vids).
contractors give:
instant scale up/down without layoff optics
no benefits overhead
no severance obligations
easy performance management (just don't renew)
this mirrors what other industries typically do after large restructuring waves ... manufacturing got temp agencies and staffing firms as permanent fixtures post-rust belt collapse. tech is just catching up to the same playbook.
clawdbot also rode the wave of claude-code being popular (perhaps due to underlying models getting better making agents more useful). a lot of "personal agents" were made in 2024 and early 2025 which seem to be before the underlying models/ecosystems were as mature.
no doubt we're still very early in this wave. i'm sure google and apple will release their offerings. they are the 800lb gorillas in all this.
very excited to see the agentic sessions when you release them.. that kind of transparency is super valuable for the community. i can see "build a browser from scratch" becoming a popular challenge as people explore the limits of agentic coding and try to figure out best practices for workflows/prompting. like the new "build a ray tracer" or say nanogtp but for agents.
it's amazing how far we've come in 20 years. i was a (very minor) contributor to khtml/konqueror (before apple got involved w/ webkit) in the early 2000s, and back then it was such a labor intensive process to even create a halfway working engine. like, months of work just to get basic rendering somewhat correct on a very small portion of the web (which was obv much smaller)
in addition to agentic coding, i think for this specific task having css-spec/html-spec/web-platform-tests as machine readable test suites helps a LOT. the agent can actually validate against real specs.
back in the day, despite having gecko as an open source reference, in practice the "standards" were whatever IE was doing. so you'd spend weeks implementing something only to discover every site was coded for IE's quirks lmao. for all of their other faults, google/apple and other contributors helped bring in discipline to that.
its basically claude with hands, and self-hosting/open source are both a combo a lot of techies like. it also has a ton of integrations.
will it be important in 6 months? i dunno. i tried it briefly, but it burns tokens like a mofo so I turned it off. im also worried about security implications.
a lot of academics aren't super technical and don't want to deal with git workflows or syncing local environments. they just want to write their paper.
overleaf lets the whole research team work together without anyone needing to learn version control or debug their local texlive installation.
also nice for quick edits from any machine without setting anything up. the "just install it locally" advice assumes everyone's comfortable with that, but plenty of researchers treat computers as appliances lol.
i'm rather unfamiliar with his work post-mono.
that interpolation is where synthesis happens. whether it is coherent or not depends.
nathan lambert (who wrote the RLHF book @ https://rlhfbook.com/ ) describes this as the "elicitation theory of post training", the idea is that RLHF is extracting and reshaping what's already latent in the base model, not adding new knowledge. as he puts it: when you use preferences to change model behavior "it doesn't mean that the model believes these things. it's just trained to prioritize these things."
so like when you RLHF a model to not give virus production info, you're not necessarily erasing those weights, the theory is that you're just making it harder for that information to surface. the knowledge is still in the compression, RLHF just changes what gets prioritized during decompression.
classical search simply retrieves, llms can synthesize as well.