Same. I switched my efforts to a larger Gemma 4 MoE model (26B-A4B) and llama.cpp and started getting meaningful results. I also implemented subagents for querying, determining which object/action to execute, and composing a short title. This is all running on an M4 in approximately 16 gb of ram. Also using Google's native tool calling channels.
There's another problem here which is that there isn't enough content. I've on multiple occasions now thrown various news perspectives into AI and asked it to research what the actual facts of a contested issue were. In most cases, it comes down to one quote from one speech. The spin was pages and pages of commentary, most of which is opinion based. The news outlet wouldn't have enough to report if they just told you the quote.
It's not like the one comment is the only newsworthy thing that day. It's a big complicated world with so many important things happening that we can't possibly keep up. It'd be easy to fill up the news with a tiny fraction of them, instead of picking one or two facts and spouting lots of opinions about them. Doing the latter is just a choice they've made.
Reporting novel facts is expensive. You need a travel budget, a widely-dispersed bureau network, a good research department, highly motivated journalists, and the list goes on. Opinions about existing facts are much cheaper to produce.
Multimodal models are trained knowing how to understand encoded images. It really is magic. Base64 image data is a binary-to-text encoding scheme that represents image bytes as a printable ASCII string, formatted as data:[<mediatype>][;base64],<data>. We think of llms as only good at text, but any structured data is predictable. As long as it can be turned into an N dimensional vector that represents a complex idea in the LLM hidden weights, the output of the model treats that essentially like text. With sufficient training data, it understands the text in what to us looks like noise.
How does security and isolation work? If someone else's account is compromised, how do I know I won't be? If instant is compromised, how do I know I won't be?
If someone else's account is compromised, you would not be, because apps are logically separated. There would be no way for the compromised or uncompromised account to ever see your data.
If Instant is compromised, then that's a lot more dangerous. We minimize this risk following security best practices: keeping data encrypted at rest, keeping secrets hashed at creation time, etc.
"logically separated" as opposed to "physically separated" (pretty rare in the Cloud world)
If you want more details, read their open source codebase or ask them specifically what documentation would boost your confidence, instead of leaving snarky comments.
I would argue that saying the accounts are logically separated is a snarky comment. It’s akin to patting the reader on the head and saying “don’t you worry your pretty little head”. Logically separated says nothing. Distinct VMs are logically separated, containers are logically separated, as are storing data in different files which self-modifying PHP code which doesn’t check its inputs tries to keep distinct. It’s basically just saying their engineers do their best but any single bug leaks data. Which is better than saying their engineers don’t even try? Not really. It’s a completely empty statement.
Also, for people who actually care about security in the cloud, physically separated is not uncommon. Side channel attacks are real. Dedicated instances are not that hard if you really care about security.
It'd be useful to understnad the nature of that logical separation: for example is data from different tenants stored on disk using different encryption keys? what about in memory? or perhaps there's no encryption-level isolation but you're relying on an authorization layer to authorize to different pieces of data: if that's the case is that built on Postgres's row-level security, for example?
These are fundamental points to be open and transparent about to instill confidence
HOW are they logically separated? Are there any layers to this security? Any standard established security boundaries like containers? Or is it just your app code doing its best not to have security bugs?
When a political party controls the science, you end up with say, Trump pushing one set of results, and Biden pushing another. It then becomes either pick the science that agrees with your politics, or throw up your hands in frustration. The average reader probably won't be able to dig into the fundamentals of the research and pull out the salient results, nor are they guaranteed it isn't policy pushed through overstated claims. It really undermines good science. It also falls back on the researchers who push science based on politics as well, so it isn't just the politicians.
That’s all true but it’s orthogonal to what i think you are responding to IMO.
I don’t think that studies are bunk because of corporate money, I think they are bunk because of how many studies I read. I am not a very fun person so when I see news reports about studies, I try to look them up. I find it more peaceful than memes or celebrity events. Think “coffee is good/bad for you again” style studies we read about daily.
These studies always suck. Ok, maybe 90% if I’m not being hyperbolic. It’s small sample sizes sure, but it’s also faulty logic, unsupported claims from evidence, lack of looking for alternatives, lack of ruling out confounding factors. And don’t get me started on soft-science an arts theses, I don’t have time.
I know that science moves forward mostly in millimeters, and I would agree that we have more and better scientific knowledge now than we have had in the past. But it certainly isn’t for the amount of publish-or-perish, p-hacking, storytelling, or outright fabrication.
In my experience the lay people aren't getting a great idea of the state of the field. Like you say coffee good/bad studies. It isn't that simple. People test all sorts of things in different contexts. But maybe some are genuinely bad studies. You don't know though because science journalism is so crappy. They don't care about the merits of the study. They go "people drink coffee, maybe that would drive engagement."
The most interesting papers are not going to get popular press releases because they are so many steps removed from the context that lay people understand. They can understand "coffee good/bad." They can't understand anything about the stories we are actually telling at the bleeding edge of a field, because even our undergrads working in our labs on these projects can scarcely understand them. Second year grad students struggle to understand them. How can a science journalist who only has a bs from communications department, or the lay public, possibly understand?
So, they don't reach for those papers when they seek to write articles for engagement. And the lay public doesn't learn the state of the art, and assumes the worst of the field from what they do read about.
Every few days I would log on. I was only 10-14 years old. It spread by word of mouth, and I just happened to have access to my dad's computer that had a modem. I'd hop on and play tradewars or similar. There were forums, mostly about hacking/pirating content. The forums were not too distant from what reddit feels like. As a young kid, it was also the only place where unfiltered information could be found, like how to make a bomb or how to get around copy protections. A lot of friends I had at the time where starting to do more serious file sharing, though the bandwidth kept that pretty limited.
What do investors want? Returns on their investment right.
So, as an investor do you throw your money blindly at a high risk endeavor that is likely to fail due to competition, or
Do you invest in setting up a limited rent seeking market that guarantees income in the future.
Unregulated free market capitalism always turns into one large bully that dominates over everyone else because one large bully that dominates over everyone else is a very effective system. Vote based governments such as democracy are a means of attempting to ensure that said government are somewhat controlled by the people and not by a king/corporations in the first place.
For instance on Matt Stoller's blog there are endless articles about how private equity is buying up medical practices, veterinary practices, cheerleading leagues, all sorts of low-risk, high-reward rollups. You also see things like the current AI bubble where there is very much an "arms race" where it seems quite likely that investors are willing to risk wasting their money because of the fear of missing out.
Some other kind of social system is going to face the same trade-offs and note that "communism" in the sense of the USSR and China might not be a true alternative. I mean, Stalin's great accomplishment was starving his peasants to promote rapid industrialization (capital formation!) so they could fight off Germany and then challenge the US for world supremacy. People who are impressed with China today are impressed that they're building huge solar farms, factories that build affordable electric cars, have entrepreneurial companies that develop video games and social media sites, etc. That is, they seem to out-capitalize us.
-ish. I often keep md files around and after a successful task. I ask Codex to write the important bits down. Then, when I come around to a similar task in the future, I have it start at the md file. It's like context that grows and is very localized. It helps when I'm going through multiple repos at multiple levels.
I’m also doing similar with fairly decent results. AGENTS.md grows after each session that resulted in worthwhile knowledge that future sessions can take advantage of. At some point I assume it will be too big, then it’s back to the Stone Age for the new agents, in order to release some context for the actual work.
