Australian startup Springboards has built an LLM called Flint that is meant to be less predictable than mainstream chatbots, according to reporting by MIT Technology Review’s Will Douglas Heaven. The pitch is aimed at a familiar annoyance: ask several big AI assistants for ideas, and they often converge on the same safe, obvious answer.
MIT Technology Review described a quick test: ask Claude, ChatGPT or Gemini for a random number from 1 to 10, and the answer is very often 7. That example is not a benchmark, and it will not reproduce every time. It is still a useful little irritant because it shows how systems sold as flexible assistants can behave more like autocomplete with a favorite chair.
Springboards’ answer, according to the report, is Flint, a model trained to produce a broader spread of replies to open-ended requests. The examples given are not narrow factual tasks. They are prompts such as asking where to travel in Europe, where a user may want options that are less recycled than the usual greatest-hits itinerary.
The distinction matters for people using chatbots as creative or planning tools. Predictability can be useful when the job is coding, research or other tasks where consistency and correctness matter more than novelty. It is a worse fit when the user is brainstorming, comparing travel ideas or trying to get out of the first answer that would occur to a search engine, a travel blog and apparently three rival chatbots.
The report does not provide independent performance results for Flint, so Springboards’ claim should be treated as exactly that: a claim about how it trained and positioned its model. The useful technical question is whether diversity in answers can be increased without turning the model into a randomness machine that is less helpful or less accurate. The available details do not settle that.
Other technology developments cited in the same roundup
- CNN reported that scientists say they have built a cell from scratch using lab-made DNA, with the resulting cell able to feed, grow and reproduce. Quanta, New Scientist and The New York Times also covered the work and its synthetic-biology implications.
- The Financial Times reported that OpenAI proposed giving the Trump administration a 5% stake, amid political pressure over AI. CNBC reported that OpenAI suggested other major US AI companies could provide similar stakes, and Bloomberg named Anthropic, Google and Meta as possible participants.
- The BBC reported that Singapore seized a $42 million mansion connected to an investigation into alleged Nvidia chip smuggling. The Financial Times reported that Supermicro’s Taiwan offices were raided days earlier as part of the probe.
- Axios reported that Anthropic’s Fable 5 returned online, with some higher-risk queries potentially sent to less powerful models. The BBC reported that access was restored after the US lifted an export ban.
- Bloomberg reported that Meta is working on a cloud infrastructure business, including selling access to models hosted on its systems and selling raw compute capacity, details also covered by CNBC and TechCrunch.
This story draws on original reporting from MIT Technology Review.
