0:12
ConlangCrafter Turns AI to Imagining Languages
There are over 7,000 natural languages today, but that doesn’t stop people from occasionally making up completely new ones. These constructed languages, or conlangs, include Dothraki, Klingon, and various Elvish languages. Now, an AI model called ConlangCrafter is also capable of generating new languages—and it is particularly good at it.
In a paper published 27 June in the Proceedings of the Association of Computer Linguists, researchers analyzed ConlangCrafter’s language generation abilities, reporting that it can develop a diverse array of novel languages that consistently abide by their rules.
In previous work, Gašper Beguš, an associate professor of linguistics at the University of California, Berkeley, showed how large language models (LLMs) can analyze languages to the same extent as most humans. In his most recent endeavour, he set out to push the language boundaries of AI models even further.
“Creating an entire language is not an easy task at all,” Beguš says, noting that some people have dedicated their careers to creating conlangs for movies, books, and video games.
But Beguš sees additional value in making AI models capable of creating truly novel languages beyond what humans could imagine. “[Models] are able to imagine or come up with things that we might not, and we can learn so much from that,” he says.
For example, ConlangCrafter can create new languages with unconventional communication systems, such as a language for a cephalopod species that uses colors and gestures instead of sounds. Of course, while this “color language” generated by ConlangCrafter isn’t truly what an octopus uses for communication, Beguš envisions these imaginary languages as a means for studying non-human centric languages in greater detail.
Beguš and the rest of the team, including Morris Alper, a postdoctoral researcher at Carnegie Mellon University and Moran Yankua, a Ph.D. student at Tel Aviv University , designed ConlangCrafter so that it can apply a wide range of linguistic rules in terms of how sounds are organized in a language (phonology), the relationship between word and sentence structure (morphosyntax), and vocabulary.
A random number generator regularly introduces variation so that every language comes out different. A built-in editing loop then reviews the result for contradictions and fixes them. Users can choose whatever mix of rules they want, or ask ConlangCrafter to make up its own rules.
“[Models] are able to imagine or come up with things that we might not, and we can learn so much from that.” —Gašper Beguš, University of California, Berkeley
“You can choose whatever flavor of language you want,” says Beguš. “You can create a mixed language between Japanese and Esperanto, for example.”
“The goal is for the languages to be creative, so they should all be different from each other,” says Alper, who specializes in multimodal machine learning and computational linguistics. “You also want them to be consistent, because a language is like a system of rules, and those rules shouldn’t contradict each other.”
To evaluate diversity, the team measured how much the generated languages differed from one another across key linguistic features such as the basic word order used in sentences. To evaluate consistency, they checked whether translations into each invented language correctly followed that language’s own rules.
They compared languages generated by ConlangCrafter to languages created by general-purpose LLMs, such as Gemini-2.5-Pro. “Our full system can be about twice as diverse and almost 70 percent more consistent than simply prompting an LLM to invent a new language,” says Alper.
David Mortensen, an assistant research professor at the Language Technologies Institute at Carnegie Mellon University who was not involved in the work, says that ConlangCrafter could help natural language processing researchers better evaluate the ways in which the structure of a language affects the performance of a model.
“There is a substantial body of research that suggests that linguistic structure–both at training time and test time–does affect model performance,” he says. “Hypotheses in this area have been very hard to evaluate, however.” He adds that a tool such as ConlangCrafter could help facilitate experiments on the effects of factors such as language typology and lexicon in a scientifically sound and reliable way.
ConlangCrafter is available for free online. Its creators note that the system is currently limited in more complex linguistic dimensions such as semantics, contextual and conversational use of language, and the visual aspects of writing.
Beguš envisions expanding upon this research to study the Sapir-Whorf hypothesis, which suggests that the way we speak influences the way we think and perceive the world. For example, this could involve running simulations of different worlds, each with its own language, exploring its impact on societies. “That’ll be a nice next step,” he says.
5:45
J.P. Morgan sees a pile of red flags in the AI market
I’ve been chewing over what J.P. Morgan just flagged, and the thing that sticks out is how thin the AI profit base has become. Out of the whole S&P 500, just 42 AI‑related firms are pulling 65‑80 % of the earnings, so the upside is riding on a very narrow set of names. That concentration alone makes the whole sector feel a lot more fragile than the headlines suggest.
What’s more, the semiconductor rally that’s been feeding AI is tracing a pattern we haven’t seen since the dot‑com era—sharp price spikes followed by a steep flattening. Leveraged chip ETFs, which were already niche, have blown up five‑fold in influence since early 2024, amplifying any swing in chip valuations. It’s a feedback loop that could magnify losses if sentiment turns.
Put together, the bank sees layers of risk: market concentration on a handful of AI players, a chip market that’s behaving like a bubble, and the extra leverage that could turn a dip into a tumble. It’s a reminder that the excitement around AI is still walking a tightrope, and a shift in investor mood could hit harder than we’d like.
7:02
Half of Claude users say AI can already handle half their work according to Anthropic survey
So I was looking at this survey of Claude users, and what's interesting is that about half of them think AI can already handle at least half of their work tasks. This is based on a pretty large group of people, almost 10,000 users. What's also notable is that a significant percentage, about a quarter of them, think that in just a year, AI will be able to cover a much larger portion of their work, somewhere between 60 to 90 percent.
It seems like the people who are earlier in their careers are more concerned about this, which makes sense. On the other hand, the users who are really heavily into using Claude are actually pretty optimistic about how this will all play out for their careers.
I think what's mechanically surprising here is how quickly people are expecting AI to take on a significant portion of their work. It's not just about automating simple tasks, but really about handling a substantial amount of what they do on a daily basis. This shift is likely to have some significant implications for how we think about work and careers in the near future.
It's also worth noting that the survey gives us a glimpse into how people are actually using AI tools like Claude, and what their expectations are for how these tools will change their work lives. This kind of insight is really valuable for understanding what's happening on the ground, and how AI is being integrated into people's daily work.
8:45
Anthropic's Fable 5 could return within days as Trump administration prepares to lift restrictions
The thing that’s shifted under the hood is the policy gate. The administration that put a hold on Anthropic’s Fable 5 back in June is now moving to pull that plug, and the paperwork is already in motion. It isn’t a brand‑new model popping up; it’s the same system waiting for the green light that was stalled by safety concerns.
What’s surprising is how quickly the reversal could happen. The decision is already in the hands of the Pentagon and the NSA, and once they sign off, the model could be back online within days rather than weeks. That timing feels almost deliberate, as if the bureaucracy is finally catching up to the technical readiness.
So, if everything lines up, you’ll see Fable 5 humming again almost immediately after the final approvals. It’s a reminder that a lot of AI availability hinges on a few signatures, not just on the model itself.
9:51
We Built a Routing Layer to Cut Our AI Costs. It Broke the Product.
I’ve been chewing on this one because it shows how a clever tweak can backfire fast. The team built a routing layer that decides, on the fly, whether a request should hit a heavyweight model or a cheaper, lighter version. By shunting the bulk of traffic to the cheaper model they slashed the inference bill by more than half, and the numbers looked great at first glance.
The snag showed up a few weeks later when users started noticing fuzzier answers. The cheap model was good enough for the easy cases, but when the router mis‑routed the tougher queries, the responses degraded. Satisfaction scores dipped, and the cost savings turned out to be a false win—quality was the hidden cost they hadn’t accounted for.
What’s interesting is how they figured it out. Instead of waiting months for churn to surface, they built a detection loop that compares output consistency across the two models in real time. When the variance crossed a threshold, the system flags the routing rule for review. It’s a kind of “early warning” that catches the Pareto trap before it eats the product.
The takeaway? Anything that routes traffic for cost reasons needs a tight feedback loop on quality. A cheap shortcut can look like a win on the balance sheet, but if the user experience slips, the whole thing unravels. The methodology they share lets you spot those mismatches in days, not months, keeping the product stable while you still hunt for savings.
