Your AI agent isn't the whole system.
It's one unit of compute inside a workflow.
Give each step only the tools it needs, and you get AI that's reliable, testable and safe in production.
4
107
You don't get an industrial revolution by scaling the artisan
avalanche.run
Gabriel Lespérance retweeted
and if, for some reason, you're not already following @rikiyatakehi @dianetc_ @jacobli99 @_matthewyang @a1zhang @souradipchakr18 @NoahZiems then you definitely should before you're late to the party :D!
Gabriel Lespérance retweeted
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.…
#1 rule of sales; follow up until they say no, or until one of you dies
8
399
Gabriel Lespérance retweeted
把前沿模型的能力沉淀成便宜、稳定的控制器,确实是个很有吸引力的方向;但我会把“蒸馏”理解成系统工程,而不只是把一次成功答案压缩成更小的模型。原文提到从零搭复杂系统会消耗大量 token,基础打好后还要不要持续拉满推理,这个问题的关键在于:哪些判断已经变成可复用的结构,哪些仍然依赖当下环境。
一个可行的分层是:前沿模型负责探索、拆解难例、提出候选策略并解释失败;把经过验证的稳定部分写成规则、检索、窄任务模型或确定性程序;运行时再由便宜组件处理高频、边界清楚的步骤。遇到低置信度、分布变化、权限敏感或不可逆动作时,明确升级给更强模型或人。这样节省的不是“智能”本身,而是避免每个常规调用都重复购买同一份推理。
不过控制器能否替代前沿推理,必须按任务分布回答。客服分类、固定格式提取和已知工具调用,可能很适合压低成本;开放式研究、目标冲突、信息缺失和长链条规划,则更容易在看似合理的局部动作中累积偏差。离线基准上的成功率也不等于线上可靠:环境会变,用户表达会变,工具会失败,低概率错误还可能集中在高损失场景。
所以我更关注一条可观测的升级路径:记录控制器的置信度、工具结果、重试与人工接管;定期抽查它与强模型或人工判断的差异;把反复出现的失败补进评估集,再决定扩大哪些能力。还要把成本核算到完整任务,而不只比较单次 token:如果便宜模型需要多轮重试、人工兜底和事后修复,账面上的低价未必代表真实效率。
DSPy 式地投入 token 构建系统,价值可能正在于把一次次隐式判断变成显式、可测试、可追踪的程序。基础设施搭好之后,当然不必永远“token maxx”;但降档应该来自证据:哪些子任务已稳定、错误代价多大、何时升级、升级后能否恢复。最理想的不是让一个小控制器假装全知,而是让系统知道自己何时足够可靠、何时必须求助。
> business can use frontier intelligence to distill a capability without purchasing frontier intelligence each time it exercises that capability.
definitely something I've also been pondering after years of using DSPy, now using tons of tokens to lay down a complex system.
once the foundations are up, does it make sense token maxx as much ?
Gabriel Lespérance retweeted
Predictably idiotic
theglobeandmail.com/business…
> business can use frontier intelligence to distill a capability without purchasing frontier intelligence each time it exercises that capability.
definitely something I've also been pondering after years of using DSPy, now using tons of tokens to lay down a complex system.
once the foundations are up, does it make sense token maxx as much ?
Gabriel Lespérance retweeted
Introducing just-in-time evals for avalanche.run
Pass any model one type (Jev) request to an agent step, and that request runs on the output and trace of that agent every single time it's ran, even in production.
All of a sudden, you don't need a dedicated evals environment, its batteries included with Avalanche, and can be on every single agent in all of your workflows.
reminder that history rhymes and that big data / etl will have it's own agentic renaissance
We should be stealing a lot more from data engineering when we build AI systems.
DAGs - does this really need to be one giant agent loop?
Lineage - which sources, artifacts, prompts, and model versions produced this?
Change detection - if a source changes, which generated artifacts are now outdated?
Incremental processing - what can we keep instead of rerunning the whole agent?
ETL - what context do we pull in, what does the model do with it, and where do we save the result?
Data engineers have been solving these problems for decades.
1
4
241
so i pretty much all but fired astra / sol 6.1
my software factory is now humming w fable as main orchestrator / opus 5.5 as lane manager.
sol is now just the subagent.
2
179
ok so.
- fable planner / chief orchestrator
- opus 5.5 medium / astra medium MoA lane orchestrator
- sol 6.1 medium subagent fast
all driven by hermes
AGI
2
247
ok so a little update.
this is clearly an AI generated post but it did point me towards keeping Astra as orchestrator.
what I did end up replacing with sol is luna.
looks like the improvements to consumption are real.
GPT-6.1 Sol: Benchmaxxed Fraud or Genuine Genius? Both Are True
Two developers. Same model. Same day.
Developer 1: "Effectively useless as an orchestrator. Runs in circles. Benchmaxxed."
Developer 2: "Opus spent hours on this bug. GPT-6.1 Sol fixed it in minutes."
One of them must be wrong, right?
Actually, no. They're both right.
Understanding why will save you weeks of frustration with this model.
The first story:
@GabLesperance tried GPT-6.1 Sol as an orchestrator, the "manager" agent that coordinates tools, sub-agents and a whole repo.
His verdict was brutal: loops, defensive programming, getting stuck on tiny details. He called it "typical Sol behavior" and said he's considering going back to Anthropic.
"Benchmaxxed" (built to ace tests rather than real work) came up more than once today.
The second story:
@marc_ibrahim hit his Codex limits and gave Opus 5.5 hours on a stubborn bug. Nothing worked.
Then he tried 6.1 Sol, and the bug was cleared in minutes.
So which is it? Genius or benchmark cheater?
What I think is actually happening:
Solving a well-defined task and managing an entire workflow are completely different skills.
A human analogy: some engineers are incredible at fixing a specific hard bug and terrible at running a team. That doesn't make them bad engineers. It makes them specialists.
The early evidence says GPT-6.1 Sol is strongest at:
→ Clear, scoped coding tickets
→ Stuck bugs with a concrete target
→ Long, iterative loops on one problem
→ Computer-use tasks
And weaker at:
→ Coordinating many moving parts
→ Knowing when to stop exploring
→ Showing restraint (@SKatalystAI found it kept asking for approval even when told not to)
That's why DeepSWE scores can be excellent while orchestration feels frustrating.
Passing a coding benchmark and managing six agents are different sports. Believe both data points.
Is "benchmaxxed" fair?
Partly. Short, cheap, correct-on-the-check answers that aren't always what you wanted is a real pattern people noticed.
But calling it fake ignores the people shipping real fixes with it today.
The truth is less viral: it's a specialist being sold as a generalist.
How to use it:
1. In a multi-agent system, try Sol as the WORKER, not the manager. Give it concrete jobs.
2. If you use it as an orchestrator, give it tight rules, clear stopping points and limits on exploration.
3. For stuck bugs, it's worth a shot, even if another model already failed.
4. Test it on YOUR workflow before trusting any hot take, including mine.
My honest take:
Most AI hot takes are true for one workflow and presented as true for everyone.
"Useless" and "genius" usually just describe which job someone gave the model.
The skill that matters now isn't picking the best model. It's knowing which job to give which model.
Where do you land? 👇
Benchmaxxed or genuinely great? Tell me the task you tried and what happened.
1
232
gpt-6.1-sol is effectively useless for me
tried to use it as an orchestrator
>> it just runs in circles in typical sol behavior: defensive programming, gets hung on unimportant details.
feels very benchmaxxed
fr i can't believe this, but im considering going back to anthropic
smh
2
4
276
Gabriel Lespérance retweeted
with ai rapidly uncovering all a priori knowledge (see anthropic math breakthroughs) gameplay will shift from knowledge generation (discovering a priori knowledge) to context capture and actuation (right to observe and act)
post-agi gameplay is as follows:
player places sensor (a priori observer) into pole position to spawncamp economically valuable context (a posteriori feed) to be actuated on by agi (summation of a priori knowledge)
moats will be:
- ability to place sensor into pole position (trust / networks)
- access to compute and capital to fund cybernetic loop (resource availability)
- steering (ability to drive cybernetic loop into economically valuable pockets of the economy)
- controlling access to reality (spawncamping) and managing reflexivity (input->output->new reality)
agreed.
the paradox is that defensible value creation must take place out-of-distribution.
if you're in-distribution, then everybody can do it.
if you're out-of-distribution, then it takes foresight, judgement, and capacity to work in the abstract
The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder
We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge
2
187














