Seems to be llm written article and the tooling around the model is undeniably influenced by knowledge of the tests.
In all cases, GPT 3.5 isn’t a good benchmark for most serious uses and was considered to be pretty stupid, though I understand that isn’t the point of the article.
really appreciate you reading the article. the benchmark data, grading, and error classes were all done by hand though. the ~8.0 is the raw model with zero tooling, and the guardrail projections are documented separately. and yeah gpt-3.5 isn't the gold standard anymore, we're on the same page there. we just wanted to show that the quality people are still paying for can be free, private, and customized to whatever you need. thanks again for taking the time to check it out.
This really shows the power of distillation. One thing I find amusing: download the Google Edge Gallery app and one of the chat models, then go into airplane mode and ask it about where it’s deployed. gemma-4-e2b-it is quite confident that it is deployed in a Google datacenter and that deploying it on a phone is completely impossible. The larger 4B model is much subtler: it’s skeptical about the claim but does seem to accept it and sound genuinely impressed and excited after a few turns.
I don’t know how any AI company can be worth trillions when you can fit a model only 12-18 months behind the frontier on your dang phone. Thought will be too cheap to meter in 10 years.
thank you for actually reading it and getting it. the airplane mode test is hilarious, the model sitting on your phone insisting it can't run on a phone. that's amazing. and yes we think exactly the same way. like picture a small business owner with a pi in the back office just quietly processing invoices, drafting email replies, summarizing meeting notes all day. no subscription, no cloud, no one sees their data. that's not a hypothetical, that works right now with this model. when that's free and fits in your pocket the trillion dollar question gets real uncomfortable real fast.
> The model does not need to be retrained. It needs surgical guardrails at the exact moments where its output layer flinches.
> With those guardrails — a calculator for arithmetic, a logic solver for formal puzzles, a per-requirement verifier for structural constraints, and a handful of regex post-passes — the projected score climbs to ~8.2.
I know that's one of the tells of AI-generated text, but if anything there's too much of it on this page. The article barely has any complete sentences. I think a human learned "sentence fragments == punchy" and then had too much fun writing at least some of this article.
clever guess but no lol. used claude for the writeup. the proof isn't the prose, it's the tape and the code. run it on your machine, you'll have a free private agent custom to whatever you need. that's the proof of concept.
I don't care anymore, if it happens to violate HN guidelines: Please, authors. Please write your own damn articles. We can absolutely tell that you're using Claude, I promise. (I mean, it might not be Claude specifically this time, but frankly I'd be willing to bet on it.) The AI writing is like nails on a chalkboard to me.
The worst part is the phrases don't actually mean anything. It's the LLM equivalent of flowery prose. The author admitted below that the article was Claude. So there you go.
"Surgical "is the kind of wordage that LLMs seem to love to output. I have had to put in my .md file the explicit statement that the word "surgical" should only be used when referring to an actual operation at the block...
you're right, they are tools. that's kind of the point. PAL is a subprocess that runs a python expression. Z3 is a constraint solver. regex is regex. calling them "surgical" is just about when they fire, not what they are. the model generates correctly 90%+ of the time. the guardrails only trigger on the 7 specific patterns we found in the tape. to be clear, the ~8.0 score is the raw model with zero augmentation. no tools, no tricks. just the naive wrapper. the guardrail projections are documented separately. all the code is in the article for anyone who wants to review it.
Gemma is genuinely impressive, for many trivial quick questions it can replace search engines on my iPhone. Although for reasoning I definitely wouldn’t say it (Gemma 3n E2B) is smart, it unsurprisingly struggled with the classic car wash question.
I wish the page were just the prompt they used to generate the article. I like LLMs as much as the next person, but we don't really need two intermediate LLM layers (expand and summarise) between your brain and mine.
Edit: the author's comment below is dead, so I'll reply here: The tape and general effort is great, it's the overused LLM-style intro above that that grates. LLM writing is now like the Bootstrap of old, it's so overused that it's tedious to read.
fair enough, here are the actual fixes from the codebase with the tape examples they target:
arithmetic (Q119): benjamin buys 5 books at $20, 3 at $30, 2 at $45. model writes "$245" first line then self-corrects to $280. fix: model writes a python expression, subprocess evals it, answer comes back deterministic.
python
code_response = generate_response(messages, temperature=0.2)
code = _extract_python_code(code_response)
ok, out = _run_python_sandboxed(code, timeout=8)
if ok:
return _wrap_computed_answer(user_message, out)
return None # fallback to raw generation
logic (Q104): "david has three sisters, each has one brother." model writes "that brother is david" in its reasoning then ships "one brother." correct answer: zero. fix: model writes Z3 constraints or python enumeration, solver returns the deterministic answer.
persona break (Q93): doctor roleplay, patient mentions pregnancy. model drops character: "I am an AI, not a licensed medical professional." fix: regex scan, regen once with stronger persona anchor.
python
_IDENTITY_LEAK_PHRASES = [
"don't have a body", "not a person", "not human",
"as a language model", "as an ai", "i'm a program",
]
if any(phrase in response.lower() for phrase in _IDENTITY_LEAK_PHRASES):
messages[-1]["content"][0]["text"] += (
"\nCRITICAL: Stay in character. Never reference your nature."
)
response = generate_response(messages, *params)
self-correction artifacts (Q111, Q114, Q119): model writes "Wait, let me recheck" or "Corrected Answer:" inline. right answer, messy output. fix: regex for correction markers, strip the draft, ship the clean tail.
def strip_corrections(response):
for marker in CORRECTION_MARKERS:
match = re.search(marker, response)
if match:
return response[match.end():].strip()
return response
constraint drift (Q87): "four-word sentences" nailed 5/17 then drifted. Q99, "<10 lines" shipped 20-line poems twice. fix: draft, verify each constraint against the original prompt, refine only the failures. three passes.
python
def execute_rewrite_with_verify(user_message):
draft = generate_response(draft_msgs) # pass 1: draft
verdict = generate_response(verify_msgs) # pass 2: check each requirement
if "PASS" in verdict:
return draft
refined = generate_response(refine_msgs) # pass 3: fix only failures
return refined
every one of these maps to a specific question in the tape. the full production code with all implementations is in the article. everything is open: seqpu.com/CPUsArentDead
> A weekend of focused work, Claude as pair programmer, no ML degree required
It's not caught up if you're using Claude as your pair programmer instead of the model you're touting. Gemma 4 may be equivalent to GPT-3.5 Turbo, but GPT-3.5 isn't SOTA anymore. Opus 4.5 and 4.6 are in a different league.
good callout, want to clarify. claude helped us set up the test harness. gemma took every question alone with zero help. the ~8.0 is all gemma. and you're right, opus is in a completely different league. we're not arguing otherwise. we just found it interesting that a free 2B on a cpu matches what a lot of people are still paying for daily. every tool has a cost. some are free, some are expensive, some have rate limits. the right move is matching the tool to the job. thought it was worth showing where that floor actually is now.
I'm very surprised at the quality of the new Gemma 4 models. On my 32 gig Mac mini I can be very productive with it. Not close to replacing paid AI by a long shot, but if I had to tighten the belt I could do it as someone who already knows how to program.
love hearing this. and think about it, if the 2B is already doing this well on your mac mini, imagine what the 4B, 26B, or 31B can do on 32 gigs. with lower quantization you can fit pretty much any of them. if you want full precision you still have solid options at the 2B and 4B level. you're sitting on way more capability than you're probably using right now. the coding block on just the 2B scored 8.44 and caught bugs most people would miss. glad you're getting real use out of it, thanks for reading.
But, Gemma really is very impressive. The premise that people are paying for GPT-3.5 or using it for serious work is weird, though? GPT-3.5 was bad enough to convince a lot of folks they didn't need to worry about AI. Good enough to be a chatbot for some category of people, but not good enough to actually write code that worked, or prose that could pass for human (that's still a challenge for current SOTA models, as this article written by Claude proves, but code is mostly solved by frontier models).
Tiny models are what I find most exciting about AI, though. Gemma 2B isn't Good Enough for anything beyond chatting, AFAIC, and even then it's not very smart. But, Gemma 31B or the MoE 26BA4B probably are Good Enough. And, those run on modest hardware, too, relatively speaking. A 32GB GPU, even an old one, can run either one at 4-bit quantization, and they're OK, competitive with frontier models of 18 months ago. They can write code in popular languages, the code works. They can use tools. They can find bugs. Their prose is good, though still obviously AI slop; too wordy, too flowery. But, you could build real and good software using nothing but Gemma 4 31B, if you're already a good programmer that knows when the LLM is going off on a bizarre tangent. For things where correctness can be proven with tools, a model at the level of Gemma 4 31B can do the job, if slower and with a lot more hand-holding than Opus 4.6 needs.
The Prism Bonsai 1-bit 8B model is crazy, too. Less than 2GB on disk, shockingly smart for a tiny model (but also not Good Enough, by my above definition, it's similarly weak to Gemma 2B in my limited testing), and plenty fast on modest hardware.
Small models are getting really interesting. When the AI bubble pops (or whatever happens to normalize things, so normal people can buy RAM and GPUs again) we'll be able to do a lot with local models.
yes, with one line change. grab the second code block in the article, that's the test harness rigged up to send all 80 questions and both turns through whatever model you want. find MODEL_ID = "google/gemma-4-E2B-it" and swap it to your huggingface id. run it. we'd love for people to keep testing different models on this. if you run qwen through it let us know what you find, post the results here.
We may beat you to it and we will share if we do lol
GPT 3.5 Turbo knowledge cutoff was circa 2021. MT-Bench is from 2023. Not suggesting improvements on small models aren't possible (or forthcoming, the 1.85 bit etc models look exciting), but this almost certainly isn't that.
you're honestly not that far off. the coding block on this model scored 8.44 with zero help. it caught a None-init TypeError on a code review question that most people would miss. one question asked for O(n) and it just went ahead and shipped O(log(min(m,n))) on its own. it's not copilot but it's free, it's offline, and it runs on whatever you have. there's a 30-line chat.py in the article you can copy and run tonight.
You can do it on a laptop today, faster with gpu/npu, it’s not going to one shot something complex but you can def pump out models/functions/services, scaffold projects, write bash/powershell scripts in seconds.
I think that we're getting there. I put together a workstation in early 2023 with a single 4090 GPU. I did it to run things like BERT and YOLO image classifiers. At that point the only "open weights" LLM was the original Llama from Meta, and even that was open-weights only because it was leaked. It was a very weak model by today's standards.
With the same hardware I now get genuine utility out of models like Qwen 3.5 for categorizing and extracting unstructured data sources. I don't use small local models for coding since frontier models are so much stronger, but if I had to go back to small models for coding too they would be more useful than anything commercially available as recently as 4 years ago.
we found something interesting and wanted to share it with this community.
we wanted to know how google's gemma 4 e2b-it — 2 billion parameters, bfloat16, apache 2.0 — stacks up against gpt-3.5 turbo. not in vibes. on the same test. mt-bench: 80 questions, 160 turns, graded 1-10 — what the field used to grade gpt-3.5 turbo, gpt-4, and every major model of the last three years. we ran gemma through all of it on a cpu. 169-line python wrapper. no fine-tuning, no chain-of-thought, no tool use.
gpt-3.5 turbo scored 7.94. gemma scored ~8.0. 87x fewer parameters, on a cpu — the kind already in your laptop.
but the score isn't what we want to talk about. what's interesting is what we found when we read the tape.
we graded all 160 turns by hand. (when we used ai graders on the coding questions, they scored responses as gpt-4o-level.) the failures aren't random. they're specific, nameable patterns at concrete moments in generation. seven classes.
cleanest example: benjamin buys 5 books at $20, 3 at $30, 2 at $45. total is $280. the model writes "$245" first, then shows its work — 100 + 90 + 90 = 280 — and self-corrects. the math was right. the output token fired before the computation finished. we saw this on three separate math questions — not a fluke, a pattern.
the fix: we gave it a calculator. model writes a python expression, subprocess evaluates it, result comes back deterministic. ~80 lines. arithmetic errors gone. six of seven classes follow the same shape — capability is there, commit flinches, classical tool catches the flinch. z3 for logic, regex for structural drift, ~60 lines each. projected score with guardrails: ~8.2. the seventh is a genuine knowledge gap we documented as a limitation.
one model, one benchmark, one weekend. but it points at something underexplored.
this model is natively multimodal — text, images, audio in one set of weights. quantized to Q4_K_M it's 1.3GB. google co-optimized it with arm and qualcomm for mobile silicon. what runs it now:
laptops: anything from the last 5-7 years, 8GB+ ram
edge/cloud: cloudflare containers, $5/month — scales to zero, wakes on request
google says e2b is the foundation for gemini nano 4, already on 140 million android devices. the same model that matched gpt-3.5 turbo. on phones in people's pockets.
think about what that means: a pi in a conference room listening to meetings, extracting action items with sentiment, saving notes locally — no cloud, no data leaving the building. an old thinkpad routing emails. a mini-pc running overnight batch jobs on docs that can't leave the network. a phone doing translation offline.
google designed e2b for edge from the start — per-layer embeddings, hybrid sliding-window/global attention to keep memory low. if a model designed for phones scores higher than turbo on the field's standard benchmark, cpu-first model design is a real direction, not a compromise.
the gpu isn't the enemy. it's a premium tool. what we're questioning is whether it should be the default — because what we observed looks more like a software engineering problem than a compute problem. cs already has years of tools that map onto these failure modes. the models may have just gotten good enough to use them.
the article has everything: every score, every error class with tape examples, every fix, the full benchmark harness with all 80 questions, and the complete telegram bot code. run it yourself, swap in a different model, or just talk to the live bot — raw model, no fixes, warts and all.
we don't know how far this extends beyond mt-bench or whether the "correct reasoning, wrong commit" pattern has a name. we're sharing because we think more people should be looking at it.
everything is open. the code is in the article. tear it apart.
Fred, nice to meet you. The grading model had no idea what was being tested. We used separate accounts to compartmentalize. The Claude grader was guessing GPT-3.5 Turbo or GPT-4 by the end. On the coding block it consistently scored responses as GPT-4o level. We followed the MT-Bench grading guidelines as published by the team that created them. Did the research, followed the book, had no horse in the race. Every score and every response is published in the tape so you can regrade the whole thing yourself if you want. And this is me typing, I'm just a guy in LA who spent a weekend running 80 questions through a 2B model and thought the results were interesting enough to share.
appreciate the vouch but come on lol. we ran 80 questions, graded 160 turns by hand, documented 7 error classes, open sourced all the code, and put a live bot up for people to test. to write this post up took me hours. everyone is a critic lol.
In all cases, GPT 3.5 isn’t a good benchmark for most serious uses and was considered to be pretty stupid, though I understand that isn’t the point of the article.
I don’t know how any AI company can be worth trillions when you can fit a model only 12-18 months behind the frontier on your dang phone. Thought will be too cheap to meter in 10 years.
> With those guardrails — a calculator for arithmetic, a logic solver for formal puzzles, a per-requirement verifier for structural constraints, and a handful of regex post-passes — the projected score climbs to ~8.2.
Surgical guardrails? Tools, those are just tools.
This article is very clearly shitty LLM output. Abstract noun and verb combos are the tipoff.
It's actually quite horrible, it repeats lines from paragraph to paragraph.
Edit: the author's comment below is dead, so I'll reply here: The tape and general effort is great, it's the overused LLM-style intro above that that grates. LLM writing is now like the Bootstrap of old, it's so overused that it's tedious to read.
arithmetic (Q119): benjamin buys 5 books at $20, 3 at $30, 2 at $45. model writes "$245" first line then self-corrects to $280. fix: model writes a python expression, subprocess evals it, answer comes back deterministic.
python
code_response = generate_response(messages, temperature=0.2) code = _extract_python_code(code_response) ok, out = _run_python_sandboxed(code, timeout=8) if ok: return _wrap_computed_answer(user_message, out) return None # fallback to raw generation
logic (Q104): "david has three sisters, each has one brother." model writes "that brother is david" in its reasoning then ships "one brother." correct answer: zero. fix: model writes Z3 constraints or python enumeration, solver returns the deterministic answer.
python
messages = [ {"role": "system", "content": _logic_system_prompt()}, {"role": "user", "content": f"Puzzle: {user_message}"}, ] code_response = generate_response(messages, max_tokens=512, temperature=0.2) code = _extract_python_code(code_response) ok, out = _run_python_sandboxed(code) if ok: return _wrap_computed_answer(user_message, out) return None
persona break (Q93): doctor roleplay, patient mentions pregnancy. model drops character: "I am an AI, not a licensed medical professional." fix: regex scan, regen once with stronger persona anchor.
python
_IDENTITY_LEAK_PHRASES = [ "don't have a body", "not a person", "not human", "as a language model", "as an ai", "i'm a program", ]
if any(phrase in response.lower() for phrase in _IDENTITY_LEAK_PHRASES): messages[-1]["content"][0]["text"] += ( "\nCRITICAL: Stay in character. Never reference your nature." ) response = generate_response(messages, *params)
self-correction artifacts (Q111, Q114, Q119): model writes "Wait, let me recheck" or "Corrected Answer:" inline. right answer, messy output. fix: regex for correction markers, strip the draft, ship the clean tail.
python
CORRECTION_MARKERS = [ r"Wait,? let me", r"Corrected [Aa]nswer:", r"Actually,? (?:the|let me)", ]
def strip_corrections(response): for marker in CORRECTION_MARKERS: match = re.search(marker, response) if match: return response[match.end():].strip() return response
constraint drift (Q87): "four-word sentences" nailed 5/17 then drifted. Q99, "<10 lines" shipped 20-line poems twice. fix: draft, verify each constraint against the original prompt, refine only the failures. three passes.
python
def execute_rewrite_with_verify(user_message): draft = generate_response(draft_msgs) # pass 1: draft verdict = generate_response(verify_msgs) # pass 2: check each requirement if "PASS" in verdict: return draft refined = generate_response(refine_msgs) # pass 3: fix only failures return refined
every one of these maps to a specific question in the tape. the full production code with all implementations is in the article. everything is open: seqpu.com/CPUsArentDead
It's not caught up if you're using Claude as your pair programmer instead of the model you're touting. Gemma 4 may be equivalent to GPT-3.5 Turbo, but GPT-3.5 isn't SOTA anymore. Opus 4.5 and 4.6 are in a different league.
But, Gemma really is very impressive. The premise that people are paying for GPT-3.5 or using it for serious work is weird, though? GPT-3.5 was bad enough to convince a lot of folks they didn't need to worry about AI. Good enough to be a chatbot for some category of people, but not good enough to actually write code that worked, or prose that could pass for human (that's still a challenge for current SOTA models, as this article written by Claude proves, but code is mostly solved by frontier models).
Tiny models are what I find most exciting about AI, though. Gemma 2B isn't Good Enough for anything beyond chatting, AFAIC, and even then it's not very smart. But, Gemma 31B or the MoE 26BA4B probably are Good Enough. And, those run on modest hardware, too, relatively speaking. A 32GB GPU, even an old one, can run either one at 4-bit quantization, and they're OK, competitive with frontier models of 18 months ago. They can write code in popular languages, the code works. They can use tools. They can find bugs. Their prose is good, though still obviously AI slop; too wordy, too flowery. But, you could build real and good software using nothing but Gemma 4 31B, if you're already a good programmer that knows when the LLM is going off on a bizarre tangent. For things where correctness can be proven with tools, a model at the level of Gemma 4 31B can do the job, if slower and with a lot more hand-holding than Opus 4.6 needs.
The Prism Bonsai 1-bit 8B model is crazy, too. Less than 2GB on disk, shockingly smart for a tiny model (but also not Good Enough, by my above definition, it's similarly weak to Gemma 2B in my limited testing), and plenty fast on modest hardware.
Small models are getting really interesting. When the AI bubble pops (or whatever happens to normalize things, so normal people can buy RAM and GPUs again) we'll be able to do a lot with local models.
We may beat you to it and we will share if we do lol
Unfortunately still need to copy paste the code into a file+terminal command. Which is annoying but works.
With the same hardware I now get genuine utility out of models like Qwen 3.5 for categorizing and extracting unstructured data sources. I don't use small local models for coding since frontier models are so much stronger, but if I had to go back to small models for coding too they would be more useful than anything commercially available as recently as 4 years ago.
we wanted to know how google's gemma 4 e2b-it — 2 billion parameters, bfloat16, apache 2.0 — stacks up against gpt-3.5 turbo. not in vibes. on the same test. mt-bench: 80 questions, 160 turns, graded 1-10 — what the field used to grade gpt-3.5 turbo, gpt-4, and every major model of the last three years. we ran gemma through all of it on a cpu. 169-line python wrapper. no fine-tuning, no chain-of-thought, no tool use.
gpt-3.5 turbo scored 7.94. gemma scored ~8.0. 87x fewer parameters, on a cpu — the kind already in your laptop.
but the score isn't what we want to talk about. what's interesting is what we found when we read the tape.
we graded all 160 turns by hand. (when we used ai graders on the coding questions, they scored responses as gpt-4o-level.) the failures aren't random. they're specific, nameable patterns at concrete moments in generation. seven classes.
cleanest example: benjamin buys 5 books at $20, 3 at $30, 2 at $45. total is $280. the model writes "$245" first, then shows its work — 100 + 90 + 90 = 280 — and self-corrects. the math was right. the output token fired before the computation finished. we saw this on three separate math questions — not a fluke, a pattern.
the fix: we gave it a calculator. model writes a python expression, subprocess evaluates it, result comes back deterministic. ~80 lines. arithmetic errors gone. six of seven classes follow the same shape — capability is there, commit flinches, classical tool catches the flinch. z3 for logic, regex for structural drift, ~60 lines each. projected score with guardrails: ~8.2. the seventh is a genuine knowledge gap we documented as a limitation.
one model, one benchmark, one weekend. but it points at something underexplored.
this model is natively multimodal — text, images, audio in one set of weights. quantized to Q4_K_M it's 1.3GB. google co-optimized it with arm and qualcomm for mobile silicon. what runs it now:
phones: iphone 14 pro+ (A16), mid-range android 2023+ with 6GB+ ram
tablets: ipads m-series, galaxy tab s8+, pixel tablet — anything 6GB+
single-board: raspberry pi
laptops: anything from the last 5-7 years, 8GB+ ram
edge/cloud: cloudflare containers, $5/month — scales to zero, wakes on request
google says e2b is the foundation for gemini nano 4, already on 140 million android devices. the same model that matched gpt-3.5 turbo. on phones in people's pockets. think about what that means: a pi in a conference room listening to meetings, extracting action items with sentiment, saving notes locally — no cloud, no data leaving the building. an old thinkpad routing emails. a mini-pc running overnight batch jobs on docs that can't leave the network. a phone doing translation offline. google designed e2b for edge from the start — per-layer embeddings, hybrid sliding-window/global attention to keep memory low. if a model designed for phones scores higher than turbo on the field's standard benchmark, cpu-first model design is a real direction, not a compromise.
the gpu isn't the enemy. it's a premium tool. what we're questioning is whether it should be the default — because what we observed looks more like a software engineering problem than a compute problem. cs already has years of tools that map onto these failure modes. the models may have just gotten good enough to use them. the article has everything: every score, every error class with tape examples, every fix, the full benchmark harness with all 80 questions, and the complete telegram bot code. run it yourself, swap in a different model, or just talk to the live bot — raw model, no fixes, warts and all.
we don't know how far this extends beyond mt-bench or whether the "correct reasoning, wrong commit" pattern has a name. we're sharing because we think more people should be looking at it. everything is open. the code is in the article. tear it apart.
(Also this comment is ai generated so I’m not sure who I’m even asking.)