Some folks like having an MCP package. I agree that gives you a bit more control. But I live dangerously and run emacs in daemon mode and tell claude to just run emacsclient to execute elisp as needed. It works amazingly well. Claude has full visibility into my editor and together we can pretty much do anything.
In fact, I increasingly think of emacs as just a live lisp machine. It's a surface that I and the agents can interact against. It acts as a document store, database, control surface, etc.
lemonade stands are staffed by 11 year olds and there can be one on every corner.
these stands do better than cerebras, which is an 11 year old deeply unprofitable company building a technology that no other company has bothered to try to even get close to replicating.
Cerebras got lucky with the ban on NVIDIA chips due to "security reasons" in the middle east, securing a massive deal with the UAE's sovereign fund. Of course, the leaders made a nice donation to the Trump crypto fund and coincidentally the ban on NVIDIA was lifted. That presented a danger for Cerebras, but they also pivoted almost entirely to an inference as a service company, maximizing tokens/sec metrics on a variety of open weight models. Then they landed the contract with OpenAI which has bought them another lifeline, and helped paper over the trail of fraud that delayed the S1 and IPO for a few years. Fascinating story indeed!
The problem with the pivot is that with only 44GB of memory, you can't run models unless you nerf them or you scale the hardware out horizontally, which takes up a lot of space, power and even more cost.
This is why their API has dumb limits on it. I'd estimate their $/tok cost around $4-5000/mtok based on hardware alone, which is insane. Sure, it is fast, but if it is heavily quanted, expensive, and can't scale, you're going to be in a world of pain.
yea this isn't "and another example just off the top of my head" - this guy co-founded seamicro then cerebras both pretty badass companies, he was a vp of marketing not ceo when the sec sued not just him but the entire executive team. trudging up 20 year old nonsense to try and punch a guy who legit deserves his success is bullshit
that's an automatic, prenegotiated sell off - calling that dumping is worse than calling the ceo a felon (at least one is technically true). are any of the big ai companies actually profitable? how has cerebras not gotten their chip right? they're the only company shipping wafer scale inference hardware, and they're shipping to the largest labs in the world. and it works - I use cerebras.ai the token rate is amazing
Sell offs don't have to happen, and $25m isn't some small amount of money.
> how has cerebras not gotten their chip right?
It doesn't scale, and won't ever be profitable. They pivoted to inference, which has the unfortunate (for them) side effect of also requiring a boat load of memory. This is why they just partnered with AMD to offload onto their chips.
it kind of is, relatively - sean might just be having a galaxy blackhole supercluster installed at home for all we know
> It doesn't scale .. They pivoted to inference
they have more options than non-wafer-scale solutions, the chips are going 3d and meanwhile they benefit from the same interconnect progress as the wholly memory-bound alternatives
> Ah, investor. Explains your responses.
physician heal thyself like your not out here tryna manifest your short
> they have more options than non-wafer-scale solutions, the chips are going 3d and meanwhile they benefit from the same interconnect progress as the wholly memory-bound alternatives
If I was a Cerebras shareholder, I would be LIVID right now. Seriously, what the hell. Insiders dumping stock: CFO, COO, accounting officer. AND NO INFO about losing GPT-6.1 Ultrafast.
Explain, it looks like samdesk is in a more narrow domain but overlapping and very related methods and business. Un-palantir-like seems like it would be a stronger editorialization.
Palantir sell consulting services ("forward deployed engineering") around graph and map database platforms, using data supplied by customers. Some of their customers ask them to make surveillance aggregation systems, others ask them to make totally banal and random enterprise software (this is why you commonly see freak-outs around hospitals buying "Palantir," for example; usually they are operating as consultants making some kind of medical records warehousing system owned by the customer).
Samdesk provide risk analytics using traditional risk analysis investigation and research, which they have now welded an "AI platform" to, like everyone does these days. They aggregate crime and protest data to make a map of "places your employees might be at risk" and sell it to companies.
Samdesk would be a Palantir _customer_ if they liked parting with their money, not a competitor; Palantir operate at a much "lower level" and don't do the data collection or research on their own.
Well they're also actually partnering with Palantir too. Claude was the model integrated into their Maven system that helped the US military determine targets in Iran (including the Minab elementary school)
Your 4D chess model of their thinking is that the risk they run of mentioning a direct competitor, that their potential customers would also be considering, is so great that the value of that comparison is dwarfed by the costs of the risk, so they don't do it. This is some next level strategic thinking.
This isn't as complicated as you're making it. Both companies market themselves as surveillance tools for oppressive regimes. Flock is similar. Do they have exactly the same business model? No. But their brand is spying on the public and protecting those in power.
> Both companies market themselves as surveillance tools for oppressive regimes.
Look at some of these "surveillance tools" that they are marketing to "oppressive regimes". Their website brags about these deeply oppressive surveillance features:
- Identifying material threats to your workforce
- Integration with Concur and your HRIS
Look at some of the "oppressive regimes" they sell to, such as the evil Amnesty International: https://imgur.com/a/U4Zef3x
We're literally commenting on an article where Samdesk flagged a public protest to Anthropic security and Anthropic employees chose to not go by it. The absolute EVIL of a public protest being avoided by people is UNSPEAKABLE. That a company would ever even consider offering information about public protests to people is OPPRESSIVE.
Again, tracking anti-AI protests to help AI executives avoid being held responsible for their company's actions is oppressive. I don't care what PR this company does to try to whitewash its image. Their actions speak louder than their propaganda.
This is not at all what Palantir OR SamDesk does. SamDesk processes and sells data (eg, has agreements with various news/social media websites etc and converts that to a nice API of events). Palantir sells software to use all sorts of data, including stuff like SamDesk (hence why I suspect plenty of Palantir customers are ALSO SamDesk customers). They're complimentary rather than competing products.
> I'm a paid Gemini subscriber via Workspace Standard accounts and yet I also only have access to 3.6.
Same and I have found it extremely annoying. I actually really like the Gemini models for question/answer stuff and reach for it before Claude (the other model family I have purchased) but it's getting long in the tooth at this point and I'm finding my Gemini usage shrinking to nearly 0.
You are right that this is hard from a prompt alone, but for a slightly stranger
reason than the obvious one. The model never sees columns. It sees tokens, and a
token can be one character or nine, so "make this line 80 wide" asks it to run a
hidden tally over pieces it cannot count by looking at them. Any slip early in a
line compounds, and there is no backspace key to reach for once it is committed.
That said, the failure is not total. A model can lean on a learned feel for line
length, pick shorter or longer synonyms to land close to the target, and rewrite
a sentence when it overshoots. It will not be perfect every time, but it lands a
lot more often than pure chance would suggest. The sampling trick you mention is
the real fix: mask each token that would push a line past the limit, and force a
newline the moment the count hits the mark. That converts a fuzzy instruction to
a hard constraint with zero training. Fine tuning helps too, but mostly sharpens
the same internal counter rather than replacing it. This reply is a small proof;
if any line here is off by one, feel free to consider your point demonstrated...
How does the merge impact you then?
reply