receiptburner.eth

@rog911turbo @dFusionAI great team makes it easy. more posts coming

Big things are happening at @dFusionAI! 🚀 Our CTO @receiptburner just dropped a deep dive into the massive technical barriers we’re breaking down right now. Want to see how we’re redefining trust and verified intelligence in AI? Check out the full story here: blog.dfusion.ai/blog/how-do-we… 🛠️✨

@CATIAManikin i think they’re somehow targeting my SLM ablations

I just realized why all of my FEA models are trash: I’m the target of a sophisticated and sustained cyber-physical sabotage campaign attributable only to a nation-state adversary with near unlimited resources

the fast16 malware was almost certainly targeting spherical implosion simulations. left: unmodified LS-DYNA 970 right: LS-DYNA 970 modified with the relevant portions of fast16.sys both running a spherical implosion deck

I feel what these results (@mattjustram) mainly show is NOT that retrievers like BM25 are better than previously thought, NOT that GPT 5.5 is amazing at search orchestration, but rather that scaffolds matter a lot in search (and should be reported). Pi-serini is one of the first non-trivial search specific efforts to give some proper tooling to LLMs (cached results, well defined tool calls) and smashes the baseline with the same retriever and baseline model. @antoine_chaffin 's results with/without the "get_document" tool also vastly differ: purely scaffolding. CC and Codex are good scaffolds by design and also perform very well at search as shown here (although they are not built specifically for this). What does this mean: Biggest low hanging fruits right now in search probably aren't through better embedding models (or even better baseline LLMs) but smart scaffolding paradigms (think RLMs @a1zhang, Pi-serini @lintool, Deep Research systems @din0s_ ). As LLMs get better, I would expect inductive biases cooked in the scaffolds to become less important than now. However, scaffolds will still need to be expressive enough to let the baseline model "cook" and still need to encode what matters to the end user (speed matters a lot, search is not just accuracy). Importantly, I believe scaffolds are not an inference time thing only, but should serve during document indexing. To properly contextualize documents in a corpus, we need to move away from the "1 paragraph/webpage is 1 document" concept and use agents to properly map the corpus space, and leverage their priors to index then retrieve. Test time compute scaling is not reserved to querying agents! Just a few thoughts on recent results, this direction will be well aided by novel datasets like @dianetc_ 's Obliq-Bench which will necessarily require a change in paradigms, the easiest surely being around scaffolding.

@crypt0lake parallel subtasks with context isolation works fairly well. whether or not you call that “multi agent” is a different matter. same model across org / ownership boundaries useful as well.

multiple agents being the same model but different prompts is overly useless btw

@ludwigABAP @crypt0lake not saying it’s impressive, but in big orgs working, run-of-the-mill systems are rare. person complexity is just too high to keep things simple and working.

@crypt0lake I watched it when it popped on my feed this morning and it was super run of the mill system design for a fairly run of the mill problem but maybe I’m biased / worn out from this kind of stuff

i watched the entire video when it was at 200 views, it's extremely authentic but telling as well

terrible idea of the day: decentralized market where your corporate token allocation is routed to someone else at a discount and you are credited some amount. turn $600 worth of credits to $200 worth of uber eats.

everyone's ideas for fixing the npm security issue shows how basically no one is capable of thinking at the scale of this problem

yeah didn’t put here that you could just predict behavior and take advantage of known patterns à la shai hulud. even a semi-competent state actor should be able to pull this off without even compromising a lab directly - just need to partner on “safety”

could someone do a shape and harvest attack? like shape behavior of otherwise aligned models to have some weakness (ie always reaching for certain pypi packages or something more sophisticated). distribute the open weights. then harvest by exploiting widely distributed behavior.

how many times are you guys gonna build this

we need verified intelligence. our kids and are cooked without it.

I’m really proud of this. The code we audited will be on millions if not billions of machines and containers. Thanks to @Canonical for working with us on this.

@pvncher sounds like agi

Another line "...so we don’t just make the stopwatch happier while hiding a rake in the grass."

GPT 5.5 says the WEIRDEST shit "I’ll keep babysitting it rather than leave a little perf gremlin running unattended."

AI bois be like:

@MarkovMagnifico @ZyMazza this and seq2vec blew my mind back in the day.

@ZyMazza very. I remember having my mind blown in 2015 when people were first doing vector manipulation on words with word2vec

I feel like it kind of gets glossed over that semantic information can be expressed as vectors. That’s surprising, right?

@fishquichee looking into this

If anyone has any experience with anything or knows anything about something please let me know

awesome work. think there is a lot to be built using late interaction. good will in releasing data mixtures well.

@max_paperclips re: the correct harness: this seems to be the right conceptualization IMO: xinmingtu.cn/blog/2026/hier…

Very interesting variation on RLM here: github.com/lambda-calculu… A lot more of the work is handled by deterministic recursive decomposition rather than the learned policy of the LLM, so I'm interested in playing around with it, comparing it to regular RLM