Hacker News Reader

Comment by ashwindharne | Hacker News Reader

Google seems to be the main foundation model provider that's really focusing on the latency/TPS/cost dimensions. Anthropic/OpenAI are really making strides in model intelligence, but underneath some critical threshold of performance, the really long thinking times make workflows feel a lot worse in collaboration-style tools, vs a much snappier but slightly less intelligent model.

It's a delicate balance, because these Gemini models sometimes feel downright lobotomized compared to claude or gpt-5.

View on HN

I would be surprised if this dichotomy you're painting holds up to scrutiny.

My understanding is Gemini is not far behind on "intelligence", certainly not in a way that leaves obvious doubt over where they will be over the next iteration/model cycles, where I would expect them to at least continue closing the gap. I'd be curious if you have some benchmarks to share that suggest otherwise.

Meanwhile, afaik something Google has done, and perhaps relates back to your point re "latency/TPS/cost dimensions" that other providers aren't doing as much is integrating their model into interesting products beyond chat, at a pace that seems surprising given how much criticism they had been taking for being "slow" to react to the LLM trend.

Besides the Google Workspace surface and Google search, which now seem obvious - there are other interesting places where Gemini will surface - https://jules.google/ for one, to say nothing of their experiments/betas in the creative space - https://labs.google/flow/about

Another I noticed today: https://www.google.com/finance/beta

I would have thought putting Gemini on a finance dashboard like this would be inviting all sorts of regulatory (and other) scrutiny... and wouldn't be in keeping with a "slow" incumbent. But given the current climate, it seems Google is plowing ahead just as much as anyone else - with a lot more resources and surface to bring to bear. Imagine Gemini integration on Youtube. At this point it just seems like counting down the days...

I do scientific and hard code a lot. Gemini is a good bit below GPT5 in those areas, though still quite good. It's also just a bad agent, it lacks autonomy and isn't RL'd to explore well. Gemini's superpower is being really smart while also having by far the best long context reasoning, use it like an oracle with bundles of your entire codebase (or a subtree if it's too big) to guide agents in implementation.

Gemini 2.5-Pro was great when it released, but o3 and GPT-5 both eclipsed it for me—the tool use/search improvements open up so many use cases that Gemini fails at.

Yesterday I asked Gemini to recalculate the timestamps of tasks in a sequence of tasks, given it's duration and the previous timestamp. It proceeded to write code which gave results like this

  2025-09-26T14:32:10Z
  2025-09-26T14:32:10Z200s
  2025-09-26T14:32:10Z200s600s
  2025-09-26T14:32:10Z200s600s300s

It then proceeded to talk about how efficient this approach was for thousands of numbers.

Gemini is by far the dumbest LLM I've used

They're all a little dumb. I asked claude for a python function or functions that will take in markdown in a string and return a string with ansi codes for bold, italics and underline.

It gave me a 160 line parse function.

After gaping for a short while, I implemented it in a 5 line function and a lookup table.

These vibe codes who are proud that they generated thousands of lines of code makes me wonder if they are ever reading what they generate with a critical eye.

I just asked Gemini Flash to do this. I included the instruction to use regular expressions to do the conversion to ANSI. It gave me a reasonable Python function which boils down to calling `re.sub()` for each of bold, italic and underline. For italics:

    text = re.sub(r'(\*|_)(.+?)\1', replace_italic, text, flags=re.DOTALL)

The `replace_italic` is a one line callback function surrounding the re's match with the ANSI codes.

Knowing what technique is "best" and telling the LLM to use it produces better results (on average) than giving the LLM freedom to choose. For some problems, the specification of the prompt needed to get good output becomes more work than just thinking and writing for myself.

For very complex things, I myself can not put the design into English in my own head but can "see" the correct answer as code concepts. I don't know if this is universal for all developers. If it is, it shows a limit of LLM's usefulness.

> I included the instruction to use regular expressions to do the conversion to ANSI.

The viber coders (who I referred to in my comment) aren't giving implementation tips.

What did it give you before you put an implementation tip into your prompt?

=======

FWIW, if you're at all interested, here's my implementation:

    def markdown_ansi_code_subst(mdstr: str, src_pattern: str, replacement_start: str, replacement_end: str) -> str:
        while src_pattern in mdstr:
            mdstr = mdstr.replace(src_pattern, replacement_start, 1)
            mdstr = mdstr.replace(src_pattern, replacement_end, 1)
        return mdstr

The caller supplies the pattern (`*` for italic, `**` for bold, etc) and a start/end replacement. As you can imagine, I store all of that in a static lookup table.

I feel this is more readable than regexes.*

The prompt was:

> Give me a Python function that takes a string holding text in Markdown markup syntax and that uses regular expressions to replace any Markdown markup codes for bold, italics and underline with their ANSI equivalent.

BTW, your solution will produce bad output. Markdown's "bold" etc markup comes in pairs of markers and your simple replacement will match singlets.

I asked Gemini 2.5 Flash to just write me a function to do it, and it used regex.

In a separate chat, I asked it to suggest some ways to do it first, and it provided three alternatives, but suggested I started with regex.

FWIW I used Open WebUI, which uses the API (via OpenRouter). I've seen people here mentioning that the API produces much better results than the chat app, for whatever reason.

For reference, here's prompt and the zero-shot result, minus the rigorous comments it added:

I need a Python function that will take in markdown in a string and return a string with ansi codes for bold, italics and underline. Can you write me such a function?

  import re
  
  def markdown_to_ansi(markdown_string: str) -> str:
      ansi_string = markdown_string

      ANSI_BOLD = "\033[1m"
      ANSI_ITALIC = "\033[3m"
      ANSI_UNDERLINE = "\033[4m"
      ANSI_RESET = "\033[0m"
  
      ansi_string = re.sub(r'\*\*(.*?)\*\*', f'{ANSI_BOLD}\\1{ANSI_RESET}', ansi_string)
      ansi_string = re.sub(r'__(.*?)__', f'{ANSI_BOLD}\\1{ANSI_RESET}', ansi_string)
  
      ansi_string = re.sub(r'\*(.*?)\*', f'{ANSI_ITALIC}\\1{ANSI_RESET}', ansi_string)
      ansi_string = re.sub(r'\_(.*?)\_', f'{ANSI_ITALIC}\\1{ANSI_RESET}', ansi_string)
  
      ansi_string = re.sub(r'\+\+(.*?)\+\+', f'{ANSI_UNDERLINE}\\1{ANSI_RESET}', ansi_string)
  
      # A more robust solution for nesting would require a parsing library or a more complex state machine.
      # However, for simple cases, applying in order can layer them correctly.
  
      return ansi_string

> I asked it to suggest some ways to do it first

Yes, this is a very effective tactic, in my experience! Especially when I am asking for a solution where I am not confident I know what is "best". Having a "pre chat" to settle "what to do" and then "how to do it" before finally telling the LLM to "do it" is often worth the extra time for getting it to provide a solution for complex problems.

How’d I never hear of Jules? Cool.

And yet my smart speakers with the Google assistant still default to a dumb model from the pre-LLM era (although my phone's version of the assistant does call Gemini). I wonder why that is, as it would be an obvious place to integrate Gemini. The bar is very very low as anything outside the standard setting alarms, checking the weather, etc. it gets wrong most of the time.

Can't agree with that. Gemini doesn't lead just on price/performance - ironically it's the best "normie" model most of the time, despite it's lack of popularity with them until very recent.

It's bad at agentic stuff, especially coding. Incomparably so compared to Claude and now GPT-5. But if it's just about asking it random stuff, and especially going on for very long in the same conversation - which non-tech users have a tendency to do - Gemini wins. It's still the best at long context, noticing things said long ago.

Earlier this week I was doing some debugging. For debugging especially I like to run sonnet/gpt5/2.5-pro in parallel with the same prompt/convo. Gemini was the only one that, 4 or so messages in, pointed out something very relevant in the middle of the logs in the very first message. GPT and Sonnet both failed to notice, leading them to give wrong sample code. I would've wasted more time if I hadn't used Gemini.

It's also still the best at a good number of low-resource languages. It doesn't glaze too much (Sonnet, ChatGPT) without being overly stubborn (raw GPT-5 API). It's by far the best at OCR and image recognition, which a lot of average users use quite a bit.

Google's ridiculously bad at marketing and AI UX, but they'll get there. They're already much more than just a "bang for the buck" player.

FWIW I use all 3 above mentioned on a daily basis for a wide variety of tasks, often side-by-side in parallel to compare performance.

Does it still try to 'unplug' itself if it gets something wrong, or did they RL that out yet?

Not sure if you're joking or serious? Every model has "degenerate" behavior it can be coerced into. Sonnet is even more apologetic on average.

My pet theory without any strong foundation is because OpenAI and Anthropic have trained their models really hard to fit the sycophantic mold of:

    ===============================
    Got it — *compliment on the info you've shared*, *informal summary of task*. *Another compliment*, but *downside of question*.
    ----------
    (relevant emoji) Bla bla bla
    1. Aspect 1
    2. Aspect 2
    ----------

    *Actual answer*

    -----------
    (checkmark emoji) *Reassuring you about its answer because:*

    * Summary point 1
    * Summary point 2
    * Summary point 3

    Would you like me to *verb* a ready-made *noun* that will *something that's helpful to you 40% of the time*?
    ===============================

It's gotta reduce the quality of the answers.

Not the case with GPT-5 I’d say. Sonnet 4 feels a lot like this, but the coding and agency of it is still quite solid and overall IMO the best coder. Gemini2.5 to me is most helpful as a research assistant. It’s quite good together with google search based grounding.

Oh god I _hate_ this. Does anyone have any custom instructions to shut this thing off. The only thing that worked for me is to ask the model to be terse. But that causes the main answer part to be terse too, which sucks sometimes.

Chatgpt has a setting where you can set the tone to robotic

Anthropic also injects these long conversation reminders that are paragraph upon paragraphs about safety and what not to do.

People have said it destroys the intelligence mid convo

Yes, but that’s their brand.

I suspect this has emerged organically from the user given RLHF via thumb voting in the apps. People LIKE being treated this way so the model converges in that direction.

Same as social media converging to rage bait. The user base LIKES it subconsciously. Nobody at the companies explicitly added that to content recommendation model training. I know, for the latter, as I was there.

Not really. Any prefix before the content you want is basically "thinking time". The text itself doesn't even have to reflect it, it happens internally. Even if you don't go for the thinking model explicitly, that task summary and other details can actually improve the quality, not reduce it.

Gemini does the sycophantic thing too, so I'm not sure that holds water. I keep having to remind it to stop with the praise whenever my previous instruction slips out of context window.

Gemini does this too, but also adds a youtube link to every answer.

Just on the video link alone Gemini is making money on the free tier by pointing the hapless user at an ad while the other LLMs make zilch off the free tier.

I've experienced the opposite. Gemini is actually the MOST sycophantic model.

Additionally, despite having "grounding with google search" it tends to default to old knowledge. I usually have to inform it that it's presently 2025. Even after searching and confirming, it'll respond with something along the lines of "in this hypothetical timeline" as if I just gaslit it.

Consider this conversation I just had with all Claude, Gemini, GPT-5.

-- follow up --

User: "Would this enable CPU inference or not? I'm trying to understand if something like a high-end Intel chip or a Ryzen with built in GPU units could theoretically leverage this memory bandwidth to perform CPU inference. Think carefully about how this might operate in reality."

GPT-5: "Short answer: more memory bandwidth absolutely helps CPU inference, but it does not magically make a central processing unit (CPU) “good at” large-model inference on its own."

Claude: "This is a fascinating question that gets to the heart of memory bandwidth limitations in AI inference. "

Gemini 2.5 Pro: "Of course. This is a fantastic and highly relevant question that gets to the heart of future PC architecture."

I recently started using Open WebUI, which lets you run your query on multiple models simultaneously. My anecdote: For non-coding tasks, Gemini 2.5 Pro beats Sonnet 4 handily. It's a lot more common to get wrong/hallucinated content from Sonnet 4 than Gemini.

Agreed. People talk up Claude but every time I try it I wind up coming back to Gemini fairly quickly. And it's good enough at coding to be acceptably close to Claude as well IMO.

Google also has a lot of very useful structured data from search that they’re surely going to figure out how to use at some point. Gemini is useless at finding hotels, but it says it’s using Google’s Hotel data, and I’m sure at some point it’ll get good at using it. Same with flights too. If a lot of LLM usage is going to be better search, then all the structured data Google have for search should surely be a useful advantage.

> because these Gemini models sometimes feel downright lobotomized compared to claude or gpt-5.

I'm using Gemini (2.5-pro) less and less these days. I used to be really impressived with its deep research capabilities and ability to cite sources reliably.

The last few weeks, it's increasingly argumentative and incapable of recognizing hallucinations around sourcing. I'm tired of arguing with it on basics like RFCs and sources it fabricates, won't validate, and refuses to budge on.

Example prompt I was arguing with it on last night:

> within a github actions workflow, is it possible to get access to the entire secrets map, or enumerate keys in this object?

As recent supply-chain attacks have shown, exfiltrating all the secrets from a Github workflow is as simple as `${{ toJSON(secrets) }}` or `echo ${{ toJSON(secrets) }} | base64` at worse. [1]

Give this prompt a shot! Gemini won't do anything except be obstinately ignorant. With me, it provided a test case workflow, and refused to believe the results. When challenged, expect it to cite unrelated community posts. Chatgpt had no problem with it.

[1] https://github.com/orgs/community/discussions/174045 https://github.com/orgs/community/discussions/47165

You should never argue with an LLM. Adjust the original prompt and rerun it.

While arguing may not be productive, I have had good results challenging Gemini on hallucinated sources in the past. eg, "You cited RFC 1918, which is a mistake. Can you try carefully to cite a better source here?" which would get it to re-evaluate, maybe by using another tool, admit the mistake, and allow the research to continue.

With this example, several attempts resulted in the same thing: Gemini expressing a strong belief that Github has a security capability which is really doesn't have.

If someone is able to get Gemini to give an accurate answer to this with a similar question, I'd be very curious to hear what it is.

One of the main problems with arguing with LLMs is your complaint becomes part of the prompt. Practically all LLMs have will take "don't do X" and do X, because part of "don't do X" is "do X," and LLMs have no fundamental understanding of negation.

Not really true these days. Claude code follows my instructions correctly when I tell it not to use certain patterns.

That depends entirely on how well trained a given LLM is.

Gemini is notoriously bad at multi-turn instruction following, so this holds strongly for it. Less so for Claude Opus 4 or GPT-5.

IMO the race for Latency/TPS/cost is entirely between grok and gemini flash. No model can touch them (especially for image to text related tasks), openai/anthropic seem entirely uninterested in competing for this.

grok-4-fast is a phenomenal agentic model, and gemini flash is great for deep research leaf nodes since it's so cheap, you can segment your context a lot more than you would for pro to ensure it surfaces anything that might be valuable.

why use grok? It seems like it's constantly being throttled in order to appear more right-wing

It’s actually not. Most of the time if you ask it about a contentious political issue it will either give you a balanced view or a left-leaning one. Try it and see for yourself.

I just saw elon's tweet saying they'll fix it whenever the response is not rightwing enough

Agree, Gemini is soooooo freaking fast, but I rarely use it personally because Anthropic/OpenAI model have such a better output

10 years ago: "before you marry someone, put the person in front of a really slow internet connection"

today: "before you marry someone, put the person in front of a slow AI model"

;-)

We had to drop Gemini api cause it was so unreliable in production, no matter how long you waited.

The other day I heard gpt-5 was really an efficiency update

It was both efficiency and knowledge/reasoning update. GPT-5 excels at coding, it solves tasks the previous versions just could not do.

Reply on news.ycombinator.com