Jonathan Mohan (@JonathanMohan) profile picture
Joined December 2008
Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)Photo by Jonathan Mohan (@JonathanMohan)
Jonathan Mohan retweeted
Hard math, especially Olympiad-style problems, is a distinctive test of LLM intelligence. But public math benchmarks increasingly seem saturated. So we dug through physical books (even ones from 1800s) and papers we couldn't find anywhere online and collected a few thousand problems we believe are unseen by today's models. We tested a few popular open-weight LLMs on a sample of these problems, scoring their performance out of 100. Kimi K3 leads, but the result that caught my eye is Qwen 3.8 27B: a 27B model scoring 57.6, ahead of DeepSeek V4 Flash and Gemini 3.7 Flash. How far can smaller models go? #llm #aiformath #posttraining #olympiadmath
4
8
1
31
2,137
Jonathan Mohan retweeted
When the dust settles, we'll have to talk about the fact that not a single vulnerability was found by a US frontier model. Instead, we're spending $10k a day on open weights models like Kimi K3 and Qwen 3.8 to find vulnerabilities in Bitcoin infrastructure. It's a disaster.
116
426
87
3,096
206,847
Jonathan Mohan retweeted
I know everyone is busy right now, but I need you to take a few minutes and write me a letter of support for .bitcoin
95
223
44
700
136,126
Jonathan Mohan retweeted
The bitcoin community didn't verify/compare the entropy of generated seeds across different HWWs. But this should become a common test done by reviewers at product release and at all software/firmware update releases. HWW vendors should facilitate these tests.
2
3
2
10
1,493
Jonathan Mohan retweeted
Have been scanning repos all day with my kimi sub and have reported a few issues to various teams. If anyone wants me to do scan their repo please reach out. My sub has lots of credits that I don't use.
15
20
1
182
12,429
Jonathan Mohan retweeted
Shame on those who spent the last 2 years distracting the community with a contrived 'spam emergency' and retarded fork nonsense. Things like covenants/vaults can enhance self-custody security, and these people obstructed their delivery to users. Hope the forkplant was worth it!
MK4s are vulenerable but are millions of times harder to steal from. MK3s can be stolen from a decent laptop, mk4,5s and Q would need much more compute. The model: - attacker has a roughly 2^20 UUID range that it knows cold cards fall in - checks a 16 button press variancer in user behavior before generating the seed You should still promptly remove funds from mk4,mk5 and Q.
37
9
8
129
20,306
Jonathan Mohan retweeted
Over a thousand bitcoin moved today in an attack. Doing emergency triage to identify the source of the issue this is how it went: @AnthropicAI - Instant Downgrade. @OpenAI - A little help. @Kimi_Moonshot - One shot. The US needs to get it shits together. @DavidSacks
trying to investigate an active security incident. Got blocked by codex. Kimi came through.
35
127
20
911
124,789
Jonathan Mohan retweeted
I found 2 private keys that were part of the 500 BTC hack, so this is independently confirmed
Confirmed. Mk2/3 vuln, I don't think mk4 is but can't be certain
10
25
4
212
118,751
AI Slop Refactoring Engineer will probably be the CS job meta of 2028.
Software maintenance and software lifecycle still matters:
2
134
Jonathan Mohan retweeted
Many Bitcoiners are upset and beefing with each other. I suspect root cause is the bear market combined with TradFi markets experiencing an AI boom. Folks want a scapegoat for disappointing returns. BTC could have been more interesting but devs were told we had to ossify...
67
16
10
316
75,500
🚨 BREAKING: The FBI has successfully extracted deleted Signal messages from a suspect's iPhone via notification storage, the place where all your notifications are stored for up to one month. Notification storage stores data from all messaging apps, it's a big flaw in iOS. But there's a way to turn it off...
434
4,442
564
23,682
6,452,580
Jonathan Mohan retweeted
Replying to @MorpheusAIs
This is a big deal. Look at these pricing numbers. $25.00 on Claude $3.20 on Morpheus. Morpheus is 87% cheaper than Clause 4.5! For GLM-5 an equivalent open weight model.
2
5
21
753
Jonathan Mohan retweeted
This could be an incredible revolution in Cosmology. The Dark Energy model of the universe, which won a Nobel Prize in 2011, may be completely wrong. The accelerating expansion instead is simply because time runs faster in the voids between galaxies. Let me explain:
644
2,811
669
22,940
4,921,837
This is actually great advice
2
12
949
Jonathan Mohan retweeted
🇪🇺@Isabel_Schnabel: The digital euro is public money issued directly by the European Central Bank. Bitcoin is different – it is a speculative asset that does not fulfil the characteristics of money. #AskECB
Do you think the Bitcoin ist better as the Euro CBDC? If not, why? #AskECB
Readers added context they thought people might want to know
The Digital euro or CBDC will not fulfill the three main characteristics of money. Because of its proposed limits (3k) it will not function as a proper store of value stlouisfed.org/education/econ…
1,133
183
368
592
566,064
Jonathan Mohan retweeted
“invested over a million dollars in developing a finance-domain focused LLM named BloombergGPT within just six months … a model (AdaptLLM-7B) costing merely $100 came out surpassing BloombergGPT in performance.” Tech is brutal lol
Replying to @rasbt
BloombergGPT was great when it came out. But just to put the rapid evolution of LLMs into perspective 📌 Bloomberg invested over a million dollars in developing a finance-domain focused Large Language Model (LLM) named BloombergGPT. And within just six months of the release of BloombergGPT, a model (AdaptLLM-7B) costing merely $100 came out surpassing BloombergGPT in performance. 📌 And now fine-tuning a 7-bn parameter model can be done for less than $100 on AWS and can be completed in 5-7 Hours. 📌 Some may contend that the BloombergGPT was designed using proprietary data. 📌 However, this proprietary data made up a negligible 0.70% of the dataset. And the rest 99.30% of the data originated from public sources accessible to everybody.
24
154
17
1,137
221,630
This is perhaps one of the most important charts on AI for 2024. It was built by the amazing researcher team at @CathieDWood’s @ARKInvest. We can see the rise of open source local models are on the path to overtake massive (and expensive) cloud based closed models.
85
385
110
1,647
862,749