bradfox2

Income vs wealth tax. These aren't comparable.

Thank you!! Goodbye manticore if this works.

Multi turn review of code written by cc reviewed by codex works pretty well. Been one of the only ways to be able to deliver larger scoped features without constant bugs. I've seen them do 10-15 rounds of fix and review until complete.

Also implemented this as a gh action, works well for sentry to gh to auto triage to fix pr.

This is as designed to gatekeep these customers. Those in control of the checklists stand to benefit.

Maybe you have not received an alert but, yes it does, and it's annoying as all hell. Dirt, sun, etc all pop an alert about degraded performance.

I answered this already for one of our investors - marked the single "decline to disclose" box and sent it back.

Copilot

It's wikipedia in the 00s all over again being preached by roughly the same age and social demographic.

No, it lets good engineers parallelize work. I can be adding a route to the backend while Cline with Sonnet 3.7 adds a button to the frontend. Boilerplate work that would take 20-30 minutes is handled by a coding agent. With Claude writing some of the backend routes with supervision, you've got a very efficient workflow. I do something like this daily in a 80k loc codebase.

The research posted demonstrates the opposite of that within the scope of sequence lengths they studied. The model has future tokens strongly represented well in advance.

This is what we do today. Have you tried it against Gemini 2.0?

This is, justifiably, very similar to nuclear reactor operators. Pay needs to reflect the working conditions to attract more people (it does for reactor operators).

Great release Daniel. Applaud the consistency you have shown.

Can you release slightly bigger quant versions? Would enjoy something that runs well on 8x32 v100 and 8x80 A100.

The cyber security gatekeepers care very little about that kind of stuff. They care only about what does not get them in trouble, and AI in many enterprises is still viewed as a cyber threat.

Hey! We're building this. It's a workflow driven automated documentation system for engineers based on your document templates. We started in energy and have expanded into engineering services. Would love your feedback.

fast-draft.ai

Please, please send me an email or dm if you are interested.

I didn't say average, and the highs are higher for many days. It sounds like you've never been in extreme heat, there's a huge difference between 89F and 110+F. More people are moving here than are leaving.

There are already different classes of licensing in the US for larger industrial vehicles.

And where I live it's 110F+ for three months of the year. No one sane is biking or walking anywhere as it's too dangerous. Not everyone lives in some temperate small European city.

Having done this for domain specific engineering paperwork that looks similar to cause analysis, it does work well at param sizes << 70B.

lora does not work though, you need full parameter training if the knowledge isnt already present in pre training set.

If it was an airplane it wouldn't be at 75ft agl, he probably wouldn't have shot at it, and if he did, the charges would have been significantly more severe. The drone was 75ft high. Pretty easy to see what it is clearly.

From a enterprise software vendor perspective, cyber checklists feel like a form of regulatory capture. Someone looking to sell something gets a standard or best practice created, added to the checklists, and everyone is forced to comply, regardless of the context.

Any exception made to this checklist is reviewed by third parties that couldn't care less, bean counters, or those technically incapable of understanding the nuance, leaving only the large providers able to compete on the playing field they manufactured.

This is the way it's done in the nuclear industry across the US for power and enrichment facilities. Operational/secure section of the plant is airgapped with hardware data diodes to let info out to engineers. Updates and data are sneaker netted in.

What defines the boundaries of internal vs external? Certainly nothing about llm weights or ops should.

Is there anything better than tmux for this purpose? I need persistence but hate the key bindings and general interface.

I'd look at generators in unregulated markets and suppliers of generation equipment.

I work at the intersection of power and AI. Buy utility stock. The bottleneck for AI expansion will not be installed GPUs, but the electricity they need to run. We are in an energy crisis and not many realize it yet. Virgina, Texas, Arizona and soon to be Ohio and Georgia are out of generation capacity and the data centers are like locusts.

I feel like this ignores the complexity of the distributed training frameworks. The challenge is in making it fast at scale.

Galactica training paper from FAIR investigated citation hallucination quite thoroughly, if you havent seen it, probably worth a look. Trained in hashes of citations were much more reliable than a natural language representation.

Very cool. My company is building a very similar tool for nuclear engineering and power applications that face similar adoption challenges for LLMs. We're also incorporating the idea of 'many-to-many' document claim validation and verification. The ux allowing high speed human verification of LLM resolved claims is what were finding most important.

Deepmind published something similar recently for claim validation and hallucination management and got excellent results.

Is it implemented?