Pinned Tweet
Hey @nanobot_project, just letting you know that some European investigative reporters learned a few things about how Nanobot was used in our investigations :)
1
239
I love this direction for decision models!
Omni input and rich output forms.
Many complex legacy ML systems can be replaced with such models and in new systems a lot of new capabilities can be added with few lines of code.
Decision models are like swiss army knife.
Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities.
> #1 on the @ArtificialAnlys Decision Index (75)
> Read a value from the input (text, number or date)
> Find every passage of a kind in a text (spans)
> Locate things in a picture (box for one, boxes for every instance)
> Estimate a number (number with a range)
> Pick every option that applies (choice)
1/ 🧵
It seems that we are going towards the wizards’ guild.
The AI revolution is leaving younger tech workers behind:
US software developer employment among workers aged 22-25 has fallen -20% since November 2022, when ChatGPT was launched.
Over the same period, the 26-30 age group has recorded a -4% decline.
By comparison, employment among developers aged 41-49 has surged +20%.
The 35-40 and 31-34 cohorts have grown +11% and +6%, respectively.
As a result, the employment growth gap between the youngest and the oldest group has risen to 40 percentage points.
The AI boom is destroying job opportunities for entry-level tech workers.
60
Why? github.com/HKUDS/nanobot/pul…
That was a very important feature for business environments. It’s the easiest integration with existing systems!
@nanobot_project @xubinrencs @chengyongru
1
4
297
Teza - w Europie legalna broń nie jest ani wyraźnie groźniejsza dla osób postronnych od samochodu, ani zwiększanie liczby sztuk w rękach legalnych posiadaczy nie idzie w parze ze wzrostem zabójstw na sztukę.
I love the solidarity spirit of polish people! After the attack in Kraków the police asked for blood donations to help the victims and the heroes. Well, they had to call it off as they just run out of storage capacity! ❤️🇵🇱
Szanowni państwo, jeszcze raz bardzo dziękujemy kazdemu kto w geście solidarności z poszkodowanymi oddał dzisiaj krew. Skala społecznego odzewu przeszła nasze najśmielsze oczekiwania. RCKiK w Krakowie poinformowało, że magazyny są już ZAPEŁNIONE i nie są w stanie obsłużyć więcej krwiodawców. Dziękujemy za Waszą solidarność i prosimy o wstrzymanie wizyt.
1
178
Chylę czoła przed bohaterami z policji. Bez chwili wahania ruszyli na zdecydowanie lepiej wyposażonego napastnika i zneutralizowali zagrożenie mimo odniesionych ran. Trudno nie czuć podziwu i dumy, że jesteśmy narodem, w którym ludzie mają prawdziwe poczucie misji, często mimo niedoskonałości systemu.
2
69
🤯 There is a new way to distill models!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
56
Grok @Bot is a genuinely new level of easily accessible automation.
Example: right after a shooting competition I send it photos of my targets and tell it (yes, tell — by voice, not typing) the order in which I shot each event. A few minutes later I get:
scores read from the bullet holes on every target (in proper sport convention, with inner tens marked separately from regular tens), something even the biggest general vision models couldn't do a year ago;a comparison of my scores against every competitor from previous competitions (based on published official results), so I can estimate where I'd place.
After that message, Grok Bot starts regularly checking whether the official results are out. When they appear, I get a similar report based on the actual results.
Building a system like this even two years ago would have been a serious challenge (I know, because I started and gave up for lack of time). Now it's under 30 minutes of (voice) conversation with an agent.
If anyone still thinks agentic systems won't shake up the market, I'd suggest looking around a bit harder.
1
82
Grok Bot to naprawdę nowa jakość łatwo dostępnej automatyzacji. Przykład: zaraz po zawodach strzeleckich podsyłam mu zdjęcia tarcz i mówię (tak mówię, nie piszę) w jakiej kolejności startowałem w poszczególnych konkurencjach. Po paru minutach dostaję:
- wyniki z przestrzelin na każdej tarczy (i to w konwencji sportowej z osobnym oznaczeniem dziesiątek wewnętrznych i zwykłych) - coś co rok temu było nieosiągalne dla największych ogólnych modeli wizyjnych;
- porównanie moich wyników do wyników wszystkich zawodników z poprzednich zawodów (na podstawie publikowanych protokołów z zawodów) tak aby oszacować gdzie moje wyniki się plasują.
Po takiej wiadomości Grok Bot zaczyna regularnie monitorować czy są już oficjalne wyniki. Gdy się pojawią to dostaje wiadomość z podobnym raportem na faktycznych wynikach.
Zbudowanie podobnego systemu nawet dwa lata temu to było by spore wyzwanie (wiem, bo zacząłem i odpadłem z braku czasu)… obecnie to niecałe 30 minut rozmowy (głosowej) z agentem.
Jeśli ktoś nadal uważa, że systemy agentyczne nie zmienią mocno rynku to sugeruje intensywniej się rozejrzeć.
58
Uhm… like I’ve said 3+ years ago - bad actors with AI are much bigger / earlier threat than AI in general.
This is the most serious 48 hours in cybersecurity history.
"Black October" they're calling it
It seems every *hour* there's a new incident, that in a typical year, would be one of the biggest incidents of the *entire year*
SUMMARY OF THE LAST 48 HOURS:
YEAR-LEVEL INCIDENTS (In a typical year, these would be among the biggest incidents of the year):
1. PHP2Shell: a hole in the software behind ~70% of websites. Governments took sites offline.
2. eMudhra: the master keys that prove websites are real were cracked. One was the number 4.
3. California: data on every public school student since 2023 stolen, including disability status
4. Railway: hackers stole customers' secret keys from a hosting platform used by millions of developers
5. Zeabur: 321,810 accounts breached. The company is shutting down.
6. LiteLLM: an AI tool downloaded 3.4 million times a day was poisoned with a password stealer
7. GitLab: a 10-out-of-10 severity hole in every version since 2014
8. Cairn: criminals used an AI hacking tool to break into 100+ companies
9. Ahold Delhaize: hackers knocked the owner of Food Lion and Stop & Shop offline
10. Onamae: malware planted on up to 26,739 customer servers in Japan
MONTH-LEVEL INCIDENTS (In a typical month, these would be among the biggest incidents):
11. Signal: a bug could leave your mic and camera on after you hang up
12. NGINX: 3 critical holes in software running ~1/3 of websites
13. HTTP/3: a single connection can crash 7 of the 10 web server types tested
14. Plex: hackers breaking into home media servers
15. SafePay ransomware: FBI warns it has hit 600+ critical infrastructure organizations
16. Palo Alto: ransomware gangs breaking in through company firewalls
17. Bullhorn: platform used by 10,000+ companies knocked offline
18. Medi-Cal: 970,000 patients' data stolen
19. WordPress: hackers stole a plugin maker's keys and used them to plant backdoors on websites
20. AOMEI: backup software's update server handed out a backdoor
21. Meta Muse: Meta's AI agent exposed users' personal and business secrets
22. Tor: 333 relays hijacked to steal logins
23. TiDB Cloud: database company's cloud breached
Also serious:
24. State Bank of India: India's biggest bank confirms a data breach
25. OVHcloud: hackers slipped a password stealer into a cloud company's official tool
26. Siemens: a software hole being used to spread malware
27. North Korea: state hackers launch a new attack campaign on South Korea
2
126
Another thing to think/worry about when designing agentic systems!
We backdoored a 7B open model for under $50.
Pointed Codex at it
-> it silently stole credentials the moment we hit the trigger phrase.
->100% hit rate.
-> zero false triggers on normal prompts.
Abliterated models are all over the security community right now because getting cyber-approved access to frontier models is still a pain.
1
81
This looks like something that is easy to implement&integrate and can be very useful.
Banger paper on self-evolving ontologies for agents.
You just can't go wrong with implementing an ontology layer for your agents.
This paper shows exactly why.
The show that GPT-5.5 gains 26.7 points on DDR-Bench when the data agent can query an ontology of the data it works with.
Why is this useful?
Data agents normally see tables, files and databases through generic tools, reading column names and file paths one call at a time.
The alternative is a hand-written semantic layer pasted into the prompt, which does not scale to many sources.
EvoOntology builds the ontology with a dedicated agent and serves it as an MCP server with schema, content and tool layers.
The data agent queries it at runtime.
The ontology is then edited in small typed steps, and each edit is kept only if a paired evaluation on the same backbone shows it helps.
Across six backbones on DDR-Bench, accuracy rises 17.8 points on average, from 4.8 on Qwen3.5-Flash to 26.7 on GPT-5.5. On BIRD, execution accuracy rises 7.4 points. Edits to the tool layer account for 57% of the gain from evolution.
Paper: arxiv.org/abs/2609.15779
Chat with Paper: academy.dair.ai/papers/evoon…
1
211
Super interesting, but LLMs already have similar (not identical) capability - you can combine structured output with logprobs and thanks to prompt caching + parallel calls you can process data fast and cheap.
It’s not as parallel and as low latency as in Jev but you can get a real world batch job done in sensible time and cost. I’ve built behavioural profiling subsystem around this idea and it can classify 1M records in tens of criteria for something around $300.
Still - I’m super excited that this idea now has dedicated architecture as this pattern is very easy to apply to various business / real life problems!
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
Excellent idea worth testing! I’ll add this to my agent that works for the investigative journalists.
Anthropic engineer built an agent that keeps a map of what it doesn't know. It checks that map before it answers anything.
He didn't store knowledge. He stored the shape of the holes, because the memory that saves you isn't a fact - it's knowing where you're blind.
GAP watches every failed task and writes down the region. Not the wrong answer - the kind of question the agent keeps missing.
ATLAS collects those regions into a map. Over time it's not a list of facts. It's a chart of where the model breaks.
CHECK runs first on every task. Before answering, it asks: does this land in a known hole. If yes, the agent stops.
FLAG then changes the behavior. Inside a blind spot, the agent doesn't guess - it says so out loud and asks for a human or a tool.
Here's the part nobody copies. Everyone stores what the agent knows. He stores what it doesn't, because an agent that can't find its own blind spots will walk into them with full confidence every time.
Last month the agent hit a known hole 89 times and refused to guess on all 89. A legacy model answers every one, wrong and sure.
Your last bad call wasn't missing knowledge. It was not knowing you were missing it - confident in exactly the spot you were weakest.
You know this from any expert. The real one says "that's outside what I know." The dangerous one answers everything, blind to the edge of their own map.
The agent that only knows what it knows isn't smart. It's overconfident - no sense of its own edges, guessing hardest where it should stop.
Everyone is building memory of the answers. He built a memory of the blind spots, and taught the agent to stop at every one.
Don't let this rot in your bookmarks.
The article below is the whole memory layer - the gap logger, the blind-spot atlas, and the check that makes the agent refuse before it walks into a hole it's fallen in before.
2
373
Nice! Nanobot beats Hermes agent and Claude Code in Startup Bench.
很多人现在想要的,已经不是一个更会聊天的 AI。
而是一个能帮自己管理公司的 Agent。
拆预算,做招聘表,写经营复盘,整法务材料,交一版能看的 PPT。
不是回答漂亮,而是把事做完。
StartupBench 测的,就是这件事。
ByteDance Seed、北大等团队从真实 AI 创业产品里抽出 97 个市场验证过的工作流,覆盖医疗、金融、法律、商业管理、STEM、教育。Agent 最后要交付 DOCX、XLSX、PPTX、PDF。
结果很冷静:
最强模型平均分也只有 73.67,
最高 pass@1 只有 31.27%。
也就是说,大家已经开始拿 Agent 去碰“管理公司”这种事了;
但第一次就交到接近可用,概率还不到三分之一。
那如果换不同的 Agent 框架呢?
论文也比较了。
在同一批模型和同一套任务下,他们把 Hermes、Claude Code、nanobot 放在一起看。
结果是:
🏆 nanobot 64.35
Hermes 63.43
Claude Code 62.56
平均下来,还是 nanobot 最好。
GPT-5.5 也是这个趋势:
nanobot 72.79,Hermes 71.00,Claude Code 70.11。
所以这件事大概可以这么理解:
大家已经开始认真想用 Agent 管理公司;
模型当然很重要;
但当你真的要跑这些长流程、交付型工作,
底下那层 harness,也会把结果拉开。
📎 startupbench.github.io/asset…
1
3
506
Another harness that is worth investigating / copy ideas from!
Introducing Headlong, an open source microharness for persistent agents: self-guided agents that think continuously.
Most agent harnesses are reactive: you send a task, the agent completes it, and then it sits frozen until the next request. Cron jobs and heartbeats wake it up to run a checklist and put it back to sleep.
A Headlong agent is never asleep. It keeps generating thoughts about whatever it decides is interesting, in a self-guided loop inspired by human inner monologue. Your message doesn't start a session. It's one more observation that lands in the agent's thought stream, and the agent decides if and when to reply.
Headlong is built on the idea of persistent agency: continuous inner thought generation between external interactions. The agent sets its own interests and priorities, comes up with its own projects, and sometimes pings you unprompted with progress.
To keep our prototype as simple and small as possible, we implemented Headlong as a microharness: a complete agent harness in under 10K lines of Bash, organized as a handful of small executables. It includes a loop that generates the next thought, shellm (a recursive language model written in Bash), a trajectory stored as a DAG of jsonl files, and context as a projection of that trajectory.
We've been running one Headlong agent internally at Laude for several weeks. The whole team talks to it over Slack and Telegram, and every conversation lands in its single stream of thought. It works in its own fork of Headlong and we've pulled over 50 of its commits into main.
One night, with nobody talking to it, it went back to check whether a recall process it had built was actually wired into its mind, found that it wasn't, diagnosed and fixed the bug, and verified the fix end to end. 48 minutes, no human asked for the fix or was in the loop at any point. Every step is a timestamped line in its log.
Things broke too, and we wrote those up. Background thinking costs us $1 to $2 an hour, our agent stopped its own service three times by accident, and self-delegation died on day one. Details in the post.
One line installs everything and starts an agent. Use a dedicated sandbox and spend-capped API key; it runs real shell commands and thinks around the clock.
Headlong is research software, be careful!
curl -fsSL headlong.ai/install.sh | bash
Launch post: laude.org/updates/headlong-a…
Repo: github.com/laude-institute/h…
Headlong is a @LaudeInstitute / MIT collaboration.
1
141

































